- **— engine Calamity system (user-approved 2026-09-02).** Optional per-battle `CalamityPolicy` (forms/chance/trigger exclusions/enabled) via `Battle.create(..., calamity_policy=...)` and `Battle.arm_calamity`; inert by default (no policy armed = pre-behavior). On curse activation, AFTER the normal activation effects (and echo spread) resolve unchanged, an armed policy may ADDITIONALLY erupt one random rollable form — additive, never a replacement. Seeded/replayable RNG (`chance`, `choice`) with 0.0/1.0 short-circuit; out-of-range chance rejected at construction (BTL-980). Cascade excluded; boss curses unrollable by construction; fail-closed everywhere; `CURSE_CALAMITY` emitted only when an eruption fires; forms are content-side data (no registry registration; preserved). Source-supported Blight Calamity form (uncapped mirror of the normal spawn); `DEFAULT_ERUPTION_CHANCE = 0.15` (protected/source-derived Arsenal value; not armed in production by default). Not invented: Cascade/Haunt/boss Calamity forms stay Unknown/ Not stated (M30-CUR-002 DEFERRED). Authoritative record: `MILESTONE_CALAMITY_SYSTEM_REPORT.md`; decision entry added to DECISIONS.md post-merge. Tests: `tests/unit/test_calamity_system.py` (14).
- **Arsenal Level-3 authorized batch + Cosmology Calamity activator.** Authorized Level-3 content delivered (`human_level3` modules: 45 abilities/ 11 artifacts/ 5 consumables/ 17 statuses) including the Cosmology passive (`ability.human.cosmology`) that arms the Calamity policy at BATTLE_STARTED at the source-stated 15%, holder-scoped (opposite-side check before any RNG draw; first-armed wins; fail-closed). Registry totals: ability 144, total 469. Delivery record: `ARSENAL_L3_BATCH_REPORT.md`. Tests: `tests/unit/test_cosmology_calamity.py` (7) + `test_l3_arsenal_batch.py`.
- **H-012 — Battle Mage affinity-curse trigger completed.** The class map is wired to the Human namespace (`HUMAN_AFFINITY_CURSES`); the M19/M25 "no Human affinity content" blocker expired with M30's 16 Human affinities + Human curses; unchanged (Divine bonds on Human beings still rejected). See `test_battle_mage_affinity_curse_triggers`.
- **M30-CUR-003 resolved — (user-approved).** H1: Hemorrhage is the ONE-SHOT production behavior (10% current HP + 10% potency loss, once per activation) — the descriptive "/turn" reading is REJECTED. M1: Madness is the self-attack through the existing damage pipeline (dodgeable, crit-eligible, DR applies) — no forced-action engine, no extra turn/action machinery (boundary stands). Decision record: DECISIONS.md . No production code change was required for the lock itself.
- **Verification (2026-09-02):** full suite at the PR #8 head `1727 passed, 0 failed, 10 errors, 2 warnings` (10 known M29 fixture errors only); focused Calamity/cosmology/arsenal suites green; Ruff/mypy/ compileall clean on the new modules. Delivered to `main` via PR #8 (merge commit `a commit`). Version stays 0.19.0 (baseline); no schema/migrations.

- **Read-only batch view of the 4.1 validation gate over the whole battle table** (docs/42 §3 pipeline stage "Data Validation"; docs/19 §12/§24 validation before learning; docs/45 §12 "investigate learning problems"; docs/42 §24 "Review datasets" at summary level). Fresh audit: knowledge export (docs/31 §26) was the only other new candidate and stays D-R005 (artifact format/scope/restore undefined); all previously-rejected candidates remain blocked. No D-R005 here — aggregation of existing deterministic verdicts by arithmetic only. Version stays 0.19.0; no schema/migrations.
- **`LearningValidationService.learning_gate_overview(include_limit=10)`** (`dio.services.learning`): frozen `LearningGateOverview`/ `LearningGateFailure` — population is ALL battles (stable id order; non-completed surface as VAL-001 failures, never hidden), each validated through the SAME `validate_battle` code path (test-proven identical verdicts); outputs pass/fail counts, a per-finding-code histogram (codes sorted), and bounded failing battles (clamp [1,25]); deterministic; writes nothing.
- **`/learning_gate [limit]`** (`dio.application.cogs.learning`, extended): admin-only ephemeral sweep surface on the reward-admin convention; empty battle table → zero-overview message. M31 locked command surface gains `learning_gate`; 4.2–4.5 surfaces preserved and re-tested.
- **Tests:** new `tests/unit/test_stage46_learning_gate_sweep.py` (9 tests). Full suite `1666 passed, 0 failed, 10 errors, 2 warnings` (10 known M29 fixture errors only; baseline 1657 + 9).

- **Read-only single-record DETAIL view completing the 4.3 summary listing** (docs/31 §27 "Inspect memories"; docs/45 §7 "View stored knowledge"; §22 developer-only tools; docs/19 §26 "Learned patterns"). Fresh audit: every other cluster candidate remains D-R005/blocked (edit = fabrication; no confidence scale; no lock schema; undefined retraining/dataset/activity-log/status content; no knowledge flow/training). No D-R005 here — verbatim transcription of one stored row. Version stays 0.19.0; no schema/migrations.
- **`LearningValidationService.get_learning_record(battle_id)`** (`dio.services.learning`): frozen `LearningRecordDetail` with the FULL stored `observed_pattern` exactly as persisted; `LRN-404` when none; deterministic newest-row pick on legacy multi-row violations; never re-runs validation, never writes (proven decoupled from battle rows).
- **`/learning_view battle_id:<id>`** (`dio.application.cogs.learning`, extended): admin-only ephemeral inspection on the reward-admin convention; the pattern renders as a JSON code block, truncated only at Discord's field limit and always labeled. M31 locked command surface gains `learning_view`; 4.2/4.3/4.4 surfaces preserved and re-tested.
- **Tests:** new `tests/unit/test_stage45_learning_record_inspection.py` (9 tests). Full suite `1657 passed, 0 failed, 10 errors, 2 warnings` (10 known M29 fixture errors only; baseline 1648 + 9).

- **The explicit, admin-only correction path over persisted learning records** — named by four specs (docs/45 §7 "Remove outdated knowledge"; docs/19 §27 "Knowledge correction/ Removing bad patterns"; docs/42 §24 "Remove entries"; docs/31 §27 "Remove incorrect knowledge"). Selected by a fresh discovery audit; classification: implementation-ready — `AILearningRecord` is leaf data with no consumers (verified in code), so deletion has no downstream semantics. No D-R005; version stays 0.19.0; no schema/migrations.
- **`LearningValidationService.remove_learning_records(battle_id)`** (`dio.services.learning`): deletes ONLY the `AILearningRecord` rows sourced to the named battle (the identifier 4.3 displays; unique per the 4.2 invariant; legacy multi-row violations fully repaired); `LRN-404` when none exists — nothing deleted, safe idempotent repeat. Battle data is NEVER touched. Frozen `LearningRemovalResult` reports the removed record ids.
- **`/learning_remove battle_id:<id>`** (`dio.application.cogs.learning`, extended): admin-only ephemeral surface on the reward-admin convention + `standard_embed`/`user_message` patterns. M31 locked command surface gains `learning_remove`; 4.2/4.3 surfaces preserved and re-tested.
- **Boundary guarantees proven by tests:** reversible per docs/42 §25 (persist → remove → persist recreates a record with an IDENTICAL frozen transcription); 4.1 gate unaffected (still passes after removal); 4.3 listing reflects removal; write boundary — only ai_learning_records changes; single-battle isolation; no automatic removal anywhere; no edit/confidence/lock (still D-R005: editing frozen evidence = fabrication; no confidence scale; no lock schema).
- **Tests:** new `tests/unit/test_stage44_learning_record_removal.py` (8 tests). the prior bundle artifact was lost too); the working tree survived complete. The 4.4 checkpoint commit was withheld per the standing rules — recovery decision pending (see root RECONSTRUCTION_STATE.md).

- **Read-only, admin-only inspection of the records 4.2 persists** (docs/19 §26 "Developer Monitoring": view learned patterns, confidence levels, data sources, learning changes; docs/45 §7/§22; docs/31 §27; docs/42 §24). Selected by a fresh discovery audit; no D-R005 needed — zero new knowledge semantics. Version stays 0.19.0; no schema/migrations.
- **`LearningValidationService.list_learning_records(limit=10)`** (`dio.services.learning`): frozen `LearningRecordListing` of stored `AILearningRecord` columns only (record id, source battle, confidence, version, created timestamp, pattern kind). Deterministic order (`created_at` DESC, ties by `source_battle_id` ASC); limit clamped to [1, 25]. Never re-runs validation, never re-derives evidence, never interprets, never writes (proven decoupled: the listing works even with the battle rows deleted).
- **`/learning_records [limit]`** (`dio.application.cogs.learning`, extended): admin-only ephemeral monitoring surface on the existing `/reward_grant` permission convention + `standard_embed`/`user_message` patterns; empty-state message; "confidence 0.0 = not yet evaluated" gloss; battle ids shown only because docs/19 §26 requires viewing data sources on an admin-only ephemeral surface. M31 locked command surface gains `learning_records`.
- **Explicitly NOT implemented:** knowledge edit/removal/locking (docs/45 §7, docs/31 §27, docs/19 §27 — destructive writes needing separately specified correction/audit semantics); confidence adjustment/formulas; pattern interpretation; dataset creation/versioning; any write; LLM/provider; self-play/RL/ prediction; Stage 5; gameplay changes.
- **Tests:** new `tests/unit/test_stage43_developer_learning_monitoring.py` (9 tests). Full suite `1640 passed, 0 failed, 10 errors, 2 warnings` (10 known M29 fixture errors only; baseline 1631 + 9).

- **The explicit, admin-only write path behind the Stage 4.1 gate.** Implemented under the user-approved bounded D-R005 contract; version stays 0.19.0; no schema/migration (`AILearningRecord` existed, unwired).
- **`LearningValidationService.persist_validated_evidence(battle_id)`** (`dio.services.learning`, extended — no redundant infrastructure): the battle MUST pass the 4.1 gate (failing battle → `LRN-422`, zero records; unknown battle → `LRN-404`); a passing battle produces exactly ONE `AILearningRecord` — idempotent (existing record returned, never duplicated; checked BEFORE re-validation so a later engine-version bump cannot orphan a recorded battle); `source_battle_id` = the battle; `confidence = 0.0` ("not yet evaluated" — no formula invented); `version = 1.0`.
- **`observed_pattern` is a frozen transcription of validated evidence only:** 4.1 validation/provenance facts (verdict, recorded + intake current ruleset/engine versions, evidence counts) + Stage 3.1 battle-review facts (summary, combat profiles, key moments, turning point, evidence-cited strengths/weaknesses). Nothing is interpreted into strategies/lessons/mistakes/recommendations. Builds and alternative/counterfactual possibilities are OMITTED (data does not exist) — proven by test. Deterministic: same battle state → same transcription.
- **`/learning_record battle_id:<id>`** (`dio.application.cogs.learning`): admin-only ephemeral surface following the existing `/reward_grant` convention (guild `administrator` or `manage_guild`); normal users rejected; M31 locked command surface gains `learning_record`.
- **Boundary held:** persistence never alters battle/replay/events/ snapshots/action_history/statistics/Stage 3 outputs; NO automatic learning exists (battle finalization writes nothing — proven by end-to-end test); no strategy discovery, interpretation, confidence calculations, recommendations, coaching, counterfactuals, self-play, RL, predictions, autonomous decisions, LLM/provider, or Stage 5 work.
- **Tests:** new `tests/unit/test_stage42_validated_evidence_persistence.py` (12 tests). Full suite `1631 passed, 0 failed, 10 errors, 2 warnings` (10 known M29 fixture errors only; baseline 1619 + 12).

- **First Stage-4 unit — the developer-controlled validation gate.** Transcribes only the spec-named check dimensions; no scores, thresholds, or heuristics invented. Version stays 0.19.0.
- **`LearningValidationService`** (`dio.services.learning`): deterministic, read-only `validate_battle(battle_id)` over the M2 artifacts (Battle, Replay, battle_state_snapshots, action_history, events). Stable finding codes in fixed order — Validity VAL-001..006 (completed state; result/winner/participant consistency; Player-row referential integrity), Completeness CMP-101..110 (replay presence + structure; action/event/snapshot evidence; turn-0 snapshot; turn coverage; duplicates; malformed combat_state), Consistency CON-201..208 (battle row vs replay results on outcome/winner/turns; row-count vs replay-list equality for actions and events; non-participant actors; out-of-range action turns; snapshot row vs combat_state turn), Provenance PROV-301..304 (recorded + current ruleset/engine versions; mismatches REPORTED and failing the gate for the current-version dataset — never mixed silently, docs/21 §25). Verdict: "pass" iff zero findings, else "fail"; frozen `LearningValidation`/`ValidationFinding`/`ValidationEvidence` result objects; `LRN-404` unknown battle.
- **Gate only — not learning:** no strategy detection, mistake classification, recommendations, confidence/anomaly formulas, pattern discovery, self-play, LLM/provider, or autonomous behavior. Writes NOTHING — `ai_knowledge`, `ai_learning_records`, `theory_strategies`, `simulations` stay unwired; Stage 4.1 produces validated evidence only (a developer-explicit persistence action awaits separately specified approval semantics, docs/45 §14/§22).
- **No cog:** no developer-command convention exists in the application (docs/45 §22 unimplemented), so the gate is service-layer only, DI-registered in `bootstrap.py`; the M31 locked public command surface is unchanged.
- **Data boundary (honest):** builds/matchups (linkage never written), per-artifact attribution, alternative decisions, and all pattern/confidence semantics remain unavailable/undefined — not validated, not invented (D-R005).
- **Tests:** new `tests/unit/test_stage41_learning_validation.py` (14 tests: pass/fail paths per dimension, provenance capture + mismatch reporting, determinism, isolation, read-only incl. zero AI-table writes, LRN-404, real finalize→validate end-to-end, DI registration without command surface). Full suite `1619 passed, 0 failed, 10 errors, 2 warnings` (10 known M29 fixture errors only). No schema/migrations; no CI changes; Stage 3.1–3.4 + M2 preserved. the working tree survived complete. Checkpoint commit withheld pending recovery decision (see root RECONSTRUCTION_STATE.md).

- **Fourth Stage-3 unit — deterministic Combat Performance Statistics.** Spec-named metrics only (docs/21 §6: damage, defense, recovery groups). Version stays 0.19.0.
- **`CombatPerformanceService`** (`dio.services.performance`): read-only career aggregation over completed battles' immutable replays + per-turn snapshots — Damage: total dealt/taken, average damage per turn; Defense: damage prevented (sum of recorded ``shield_absorbed`` on damage taken), shields created (count and total value of ``shield`` effects sourced), damage avoided (dodge count + dodge rate from ``after_attack`` events); Recovery: healing done (sourced healing effects) and received, resources recovered (``energy_gained`` event amounts), survival duration (average turns survived from snapshot ``alive`` traces).
- **`/performance` command** (`dio.application.cogs.performance`): public embed on the existing `standard_embed`/`user_message`/ lookup-only-identity architecture (STX-404 unknown player). DI-registered; the M31 locked command-surface list gains `performance`.
- **Honest data boundary:** docs/21 §6 "damage efficiency" has no defined semantics in the surviving specification and is NOT computed; matchup/build analytics, artifact analysis, coaching, and counter suggestions remain data-blocked or recommendation-based and out of scope. Nothing fabricated (D-R005).
- **Explicit non-goals:** no LLM/provider, no learning, no writes, no recommendations, no Stage 4+ features, no schema/balance/reward changes.
- **Tests:** new `tests/unit/test_stage34_combat_performance.py` (7 tests): damage/defense/recovery metric math, completed-battle scoping + player isolation, empty state + zero-division safety, STX-404, read-only guarantee, embed contract, command registration; M31 surface list updated. ---

- **Third Stage-3 unit — deterministic Action Statistics.** Spec-named metrics only (docs/21 §7: ultimate timing, item usage, rest usage, target selection; ability usage shipped as Stage 3.2). Version stays 0.19.0.
- **`ActionStatisticsService`** (`dio.services.actions`): read-only aggregation over `action_history` rows of COMPLETED battles: per-category usage (attack/ability/ultimate/item/rest — uses, battles, average/first/last turn, giving docs/21 §7 "Ultimate timing" and "Rest usage"), item usage breakdown per item (`action_data.content_id`), and target selection (per-opponent targeted-action counts, distinct targets; self-targeted actions excluded). Deterministic ordering throughout (count desc, then id).
- **`/actions` command** (`dio.application.cogs.actions`): public embed on the existing `standard_embed`/`user_message`/ lookup-only-identity architecture; display names only — raw content/player ids never rendered (STX-404 unknown player). DI-registered; the M31 locked command-surface list gains `actions`.
- **Honest data boundary:** docs/21 §8 Decision Quality (counterfactual comparisons), docs/29 §12 Artifact Analysis (per-artifact attribution is not recorded — artifacts act via hooks and recorded effects carry combatant ids), docs/29 §13 Counter Suggestions, and build-level analytics (unwritten build linkage) are NOT implemented. Nothing fabricated.
- **Explicit non-goals:** no LLM/provider, no learning, no writes, no recommendations, no Stage 4+ features, no schema/balance/reward changes.
- **Tests:** new `tests/unit/test_stage33_action_statistics.py` (9 tests): category aggregation + ultimate timing, item breakdown, target selection + deterministic order, completed-battle scoping + isolation, empty state, read-only guarantee, STX-404, embed name contract, command registration; M31 surface list updated. ---

- **Second Stage-3 unit — deterministic Ability Usage Analysis.** Spec-defined metrics only (which abilities, how often, when, effectiveness, who uses them); no invented analytics. Version stays 0.19.0.
- **`AbilityUsageService`** (`dio.services.usage`): read-only aggregation over `action_history` rows of COMPLETED battles: per-player per-ability `uses`, `battles`, `average_turn`, `success_rate` (share of uses whose action resolved ok with at least one successful effect), `average_damage`/`average_healing` (from recorded per-effect results), and the deterministic most-co-used partner ability (ties -> lexicographically smallest id; docs/21 §10 "common combinations"); plus the ability-scoped aggregate across players (total uses, unique users, averages; docs/21 §10 "who uses it"). Entries ordered by uses desc, then id.
- **`/usage` command** (`dio.application.cogs.review`-style cog at `dio.application.cogs.usage`): public embed on the existing `standard_embed`/`user_message` architecture; no argument = the invoker's own summary (lookup-only identity, STX-404); `ability:<name>` = name-resolved aggregate via `ContentService.search` (USG-404 when unmatched or unrecorded). Display names only — raw content/player ids are never rendered. DI-registered; the M31 locked command-surface list gains `usage`.
- **Honest data boundary:** docs/21 §10 "Context" (player skill, build type, opponent type) is NOT implemented — build linkage (`battles.build_a_id/build_b_id`) is never written by any current code path and skill labels are judgments; docs/29 §10 Build Coaching remains data-blocked on that linkage. Nothing was fabricated.
- **Explicit non-goals:** no LLM/provider, no learning, no writes, no recommendations/coaching, no counter suggestions, no build/opponent analytics, no schema/balance/reward changes.
- **Tests:** new `tests/unit/test_stage32_ability_usage.py` (9 tests): aggregation math, partner tie-break, completed-battle scoping + isolation, cross-player aggregate, empty state, read-only guarantee, STX-404/USG-404, embed name contract, command registration; M31 surface list updated. ---

- **First Stage-3 (Basic Intelligence) unit — deterministic Battle Review. AI analyzes; it never decides.** Version stays 0.19.0 (post-crash convention; see M2 entry).
- **`BattleReviewService`** (`dio.services.review`): read-only, deterministic post-battle analysis over the completed M2 data foundation (Battle row + immutable Replay JSON + per-turn battle_state_snapshots). Outputs frozen dataclasses: summary, per-player `CombatProfile` (damage dealt/taken, healing received, crits, dodges/incoming attacks, action mix, ultimate timing), `KeyMoment`s, one turning point, and evidence-cited strengths/weaknesses (`PlayerReview`). No LLM/provider, no learning, no writes of any kind.
- **Locked deterministic rules** (operator-approved as NEW DESIGN): R1 biggest single-turn HP swing (total HP loss across both combatants, earliest on ties); R2 durable winner HP-trace lead (earliest turn from which the winner stays strictly ahead in every later snapshot); R3 elimination turn (earliest snapshot where a combatant is no longer alive); R4 ultimate timing (first ultimate turn labelled early/mid/late against total turns); turning-point priority R3 > R2 > R1; fixed factual strength/weakness mappings (out-damage, dodge/crit counts, damage imbalance, unused ultimate, decisive-swing loss share), each citing its evidence.
- **`/review` command** (`dio.application.cogs.review`): public embed via the existing `standard_embed`/`user_message` architecture; no argument = the invoker's latest completed battle; explicit battle id allowed only for battles the invoker participated in (REV-403 otherwise; REV-404 unknown battle/replay; REV-405 not completed; STX-404 unresolvable player). No internal ids are ever rendered (usernames + turn numbers only). Registered in the DI container; the M31 locked command-surface list gains `review`.
- **Explicit non-goals (unchanged):** no counterfactuals/alternatives, mistake classification, coaching, build analysis, counter suggestions, predictions, adaptive learning, autonomous decisions, balance/reward changes, schema changes.
- **Tests:** new `tests/unit/test_stage31_battle_review.py` (11 tests): summaries (victory/draw), profile aggregation, R1-R4 incl. negatives, turning-point priority, evidence-cited reviews, read-only guarantee (row counts unchanged), battle isolation + validation errors, real finalize→review end-to-end, embed id-hiding contract, command registration.
- Reconstructed/completed during post-crash recovery from the M31.0/ v0.19.0 baseline (new design, operator-approved rule set). ---

- **MINOR bump:** shared Discord presentation foundation for M31 (screenshots-as-UX-reference program). No combat mechanics, content, engine, or balance changes; no new/deferred content implemented.
- **New presentation layer:** `dio.application.presentation` (embed color tokens per surface, button-style rules — red is exit/destructive only, 10-segment meter bars, compact `5.73M` number formatting, pagination footer text, friendly restriction wording), `dio.application.names` (`NameResolver` — the single id→display-name path; never raises, never renders raw ids), `dio.application.errors` (`user_message`/`sanitize_text`/`problem_lines` — B6: clean, id-free player-facing error copy; internals stay in logs).
- **Service name/search API:** `ContentService.display_name` + `ContentService.search` (+`ContentMatch`) — registry-backed, used by all Discord surfaces; prepares autocomplete for M31.1+.
- **ID exposure cleared (M31 audit §IV):** battle embed title is `⚔️ {A} vs {B}` (was `Battle {id[:8]}`); result/feed footers drop internal ids; battle log renders `attacked`/ `used {content name}` from recorded `data.content_id` (was `used Attack`); status/opponent curse + cooldown lines resolve to display names with the battle's curse threshold (`Purgatory: 63/200`); `/info` lists are name-first; profile/character/build success echoes no longer print row ids.
- **B3 fix:** an empty `/build` invocation is now a name-based view of the current build — it can never wipe the saved loadout (full editor arrives in M31.1).
- **B5 fix (verified gap):** slash commands were never registered — no `tree.sync`/`setup_hook` existed. Startup now syncs after cog load, gated by `an internal setting` (default true) with an optional `an internal setting` fast path; failure is logged, never crashes startup.
- **Recovery id-free:** `/recover`/`/recover_end` no longer print or require internal battle ids; `/recover_end` defaults to the caller's own stale session.
- **Version pins:** 0.18.0 → 0.19.0 (`pyproject`, `dio.__version__`, version-pin tests updated per convention; M22.5 identity contracts unchanged).
- **Tests:** 52 new (`test_m31_presentation/names/errors/embeds/ foundation`); M11/M28 suites re-pointed at intentional string changes (titles/log wording/info lists/profile lead) with every covered behavior re-asserted; engine/content suites untouched. ---

- **Intake/audit milestone; version stays 0.15.2** (M29 §22). Verdict: **C — PARTIALLY READY/ SPECIFIC BLOCKERS**.
- The complete Human source package (, 373 entries; 319 at full mechanics) was discovered, read, and machine-inventoried (`scripts/m29_source_intake.py`): 16+6 curses, 16+6 ultimates, 22 consumables, 43 wildcards, 120+30 abilities, 102+12 artifacts.
- Affinity → Curse → Ultimate mapping derived from the source and machine-verified (22/22 pairs; 16 player + 6 boss). Rune reconciliation record produced for all 16 Runes (Ember→Scorch … Sanguine→Hemorrhage); bindings unchanged in M29; Ruin/Physical premise discrepancies (/) resolved by the source.
- Specific blockers: 28 ability + 4 artifact name/ID collisions with the Divine registry (race-namespace decision REQUIRED); levels-based model vs the no-levels boundary; no Affinity section in the source; 54 boss entries names-only; 6 source-ambiguous consumables; gang/prestige/shard/gemstone/stardust systems referenced but unauthorized. Nothing implemented, invented, rebalanced, renamed, or converted; Divine untouched; registry 269; no economy/Wildcard/ rating/Stardust/progression; no deployment.
- Tests: `tests/unit/test_m29_human_content_intake.py` (16 tests: source integrity, mapping derivation, collision pins, Rune reconciliation record, Divine regression, boundaries); **1123 total, 83.27% coverage**, mypy (118 files), Ruff (219 files), ready 0, fuzz 0, secret scan 0.

- **MINOR bump:** production Human content implemented on top of the M30 foundation (affinities/curses/Runes).
- **Human Ultimates (13):** Absolution, Glaciate, Largesse, Lucidity, Miasma, Bulwark, Torment, Rupture, Unmake, Accretion, Phantasm, Bloom, Celerity — source max-level values, L1→L9 harmony tables preserved (level adapter), affinity-gated; Daybreak/Enigma/Awaken deferred (documented).
- **Human Consumables (9):** Focus Crystal, Fuel Cell, Leech Seed, Lucky Voucher, Medkit, Purging Stone, Repeater, Teddy Bear, War Horn; gang/prestige items deferred.
- **Human Abilities (21)** and **Human Artifacts (20)** with real mechanics (battle-start passives, crit reactions, lifesteal, execution, grit, curse reflects, control).
- **Name-collision policy:** COLLISION_MAP (32 thematic renames; source_name provenance preserved); Divine originals untouched; Shattering Charge + Desperate Pact implemented under Human ids.
- **Generic engine extensions:** harmony pool ceiling scales to the participants' ultimate costs (Divine battles keep 100); energy-gain and curse-buildup events carry effect metadata (Torrent rest-block, Unmake spread guard); level adapter for L1→L9 tables.
- Tests: `tests/unit/test_m30_full_human_content.py` (50 tests: registration, ultimate execution, consumables, abilities, artifacts, collisions, serialization/replay, loadout isolation, Divine regression, boundaries). **1210 total, 81.60% coverage**, mypy strict clean (129 files), Ruff clean (232 files), build `andy_engine-0.17.0`, registry 389, `dio ready` 0, fuzz 0, secret scan 0. Divine unchanged; no economy/Wildcard/rating/Stardust/ progression; no deployment.

- **MINOR bump:** new production Human content.
- **Human Affinities (16):** `affinity.human.<name>` (Ember…Sanguine), identity-only (the supplied source establishes no affinity mechanics), `content_race="human"`, owning their Human curse via metadata; `HUMAN_AFFINITY_CURSES` (16 machine-verified pairs) is the authoritative Human mapping; the Divine `AFFINITY_CURSES` untouched.
- **Human Curses (16):** `curse.human.<name>` (Scorch, Freeze, Shock, Doom, Drown, Exhaust, Hex, Madness, Sow, Wither, Break, Cascade, Crush, Blight, Haunt, Hemorrhage) + 8 Human curse statuses. 14 implemented from the source via existing engine channels (status ticks/hooks, stat modifiers, action restrictions, roster ops, summons, last-action replay). Blocked & documented: Break (ability levels) and Cascade's Calamity amplification (no Calamity system). Source ambiguities recorded (Hemorrhage "/turn", Blight explosion magnitude, Madness forced-action adapter, Scorch explosion timing). Boss curses NOT registered.
- **Rune reconciliation:** all 16 Rune bindings point at the Human identities (Ember→Scorch … Sanguine→Hemorrhage); loader validates against HUMAN_AFFINITY_CURSES; Light/Verdant/Boreal/Entropic Rune effects re-keyed to Exhaust/Sow/Freeze/Cascade; the Sanguine Rune's v1 amplifier is DEFERRED (M30-RUN-001); RuneMaster/raid/eligibility layers intact.
- **Engine hardening:** EventBus.publish snapshot iteration (a status hook subscribed mid-dispatch could re-run earlier hooks on one event — found via the Cascade tests).
- Tests: `tests/unit/test_m30_human_affinity_curse.py` (37 tests); historical audit tests updated with documented justification (M22.5, M23, M24, M25, M26, M27, M28, M29, production-content, canonical rules, registry, services). **1160 total, 82.81% coverage**, mypy strict clean (121 files), Ruff clean (223 files), build `andy_engine-0.16.0`, registry 309, `dio ready` 0, fuzz 0, secret scan 0. Divine unchanged; no economy/Wildcard/rating/Stardust/ progression; no deployment.

- **PATCH bump:** Discord command/API surface changed per the established convention (§37) — gameplay-neutral, no combat formulas, balance values, Rune, economy, Wildcard, rating, or progression change.
- **Command consolidation:** the six per-type info commands (`/abilities`, `/artifacts`, `/affinities`, `/curses`, `/ultimates`, `/consumables`) were REMOVED and consolidated into `/info` — one command, eight selectable categories (Abilities · Artifacts · Consumables · Affinity & Ultimate [combined intentionally] · Curses · Runes · Race · Classes) with select-menu navigation, pagination (25/page), and registry-grounded entry info (incl. Rune bonded/ unbonded/eligibility/raid fields and class tier/slot data).
- **Battle UX (BattleView):** eight primary controls — Surrender, Attack, Rest, Ultimate, Status buttons + Ability/Artifact/Consumable selects + Target select with deterministic ⚡ Weakest/ 💪 Strongest shortcuts (HP-ordered, tie-break participant order; engine-validated).
- **Battle Status embed:** HP bar + current/max/%, Power, Energy, Harmony, Dodge, Critical, Defense; Curse Resistance shown ONLY when buildup exists; Curse Buildup, active statuses, cooldowns, restrictions, summons, Enemies count. Opponent inspection via Status → Select Opponent (observable stats + visible buildup; ability inventory hidden).
- **Action flow:** player action-result embed (authoritative engine data: damage/crit/heal/status/curse/resource effects + HP deltas), compact Battle Log in the battle embed (last 5 actions), surrender flow (`BattleService.surrender` → engine `_end_battle` outcome "surrender"), character icon integration point with safe fallback.
- **Fixed (M28-UX-001):** latent M11 embed bug (`ResourcePool.maximum` → `specs[name].max_value`); presentation-only.
- Tests: `tests/unit/test_m28_ux.py` (22 tests) + consolidated-cog tests updated; **1107 total, 83.27% coverage**, mypy strict clean (118 files), Ruff clean (218 files), build `andy_engine-0.15.2`, registry 269 unchanged, `dio ready` 0, fuzz 0, replay/isolation/ Discord green, secret scan 0. No deployment. Human content boundary and gameplay boundary preserved (§34/§29).

- **Audit-only milestone; version stays 0.15.1** (M27 §23). Verdict: **B — FRAMEWORK READY WITH GAPS**.
- Framework verified: race-typed architecture end-to-end (every content type carries `content_race`); registries race-agnostic (globally unique IDs, race metadata); loader rejects every invalid race combination (full matrix tested); resolver action paths race-agnostic; serialization/replay/battle-state/Discord embeds preserve race. Human classes (25) + Runes (16) verified intact; no Human content invented; no placeholders; no Rune binding changed; no Divine content modified.
- Gaps (pending Human source): Human Affinity→Curse mapping structure + loader path (M26-HUM-003); Rune reconciliation (/); per-curse capability classification; Human Ultimate affinity-gate applicability.
- Tests: 19 new audit tests (`tests/unit/test_m27_framework_readiness.py`); **1084 total, 83.36% coverage**, mypy/Ruff clean, registry 269, `dio ready` 0, fuzz 0, secret scan 0, Stardust 0. No deployment.

- **Audit-only milestone; version stays 0.15.1** (M26 §21).
- Finding: the Human Affinity/Curse source was NOT supplied — exhaustive workspace search found only the original prototype (the Divine source, already ported). 0 Human affinities, 0 Human curses registered (machine-verified). Nothing invented.
- All 16 Rune bindings remain UNRESOLVED — no binding changed. `AFFINITY_CURSES` remains Divine-only, untouched.
- Engine capability analysis + per-curse audit templates documented; Divine side verified unchanged; economy/Wildcard/rating/Stardust boundaries re-verified.
- Tests: 13 new audit tests (`tests/unit/test_m26_human_affinity_curse_audit.py`); **1065 total, 83.36% coverage**, mypy/Ruff clean, registry 269, `dio ready` 0, fuzz 0. Verdict: **C — AUDIT BLOCKED** (Human source incomplete). No deployment.

- **Audit-only milestone; version stays 0.15.1** (M25 §22 — no gameplay release for documentation).
- Audit: machine-generated Human/Divine content inventory (`scripts/m25_content_inventory.py`): Divine 177 (77 abilities/ 16 affinities/ 40 artifacts/ 16 curses/ 16 ultimates/ 12 consumables), Human 41 (25 classes/ 16 Runes), shared 51 statuses, registry 269.
- Source-of-truth finding: Rune `bound_affinity`/`bound_curse` resolve to Divine-registered identities; `AFFINITY_CURSES` is not accepted as the final Human source of truth until the Human Affinity/Curse spec is supplied and reconciled. Nothing renamed or rewritten.
- Missing Human systems documented as UNKNOWN (Affinities, Curses, Abilities, Artifacts beyond Runes, Consumables, Ultimates, specs); integration contract + upload checklist produced; economy/Wildcard/ rating/Stardust boundaries verified (0 Stardust references; no new systems).
- Tests: 20 new audit tests (`tests/unit/test_m25_human_content_separation.py`); **1052 total, 83.36% coverage**, mypy/Ruff clean, registry 269, `dio ready` 0, fuzz 0. Verdict: **B — HUMAN INTEGRATION PARTIALLY READY**. No deployment.

### Milestone 16 — decision resolution & final combat specification audit (2026-08-10)
- **No production code changes** (version stays 0.12.0) — the M16 brief granted no approvals of the M15 decision package; the milestone is an audit/verification/documentation lock.
- Added `tests/unit/test_m16_audit.py` (24 tests): curse audit dimensions (buildup/resistance/threshold/duration/removal/ ownership/inflictor/summons/nullification/isolation/replay), ultimate audit (canonical V1/affinity/Harmony tiers/one-action/ summon exclusion/replay), consumable audit (12 registered; Ability Cache inert and blocked), security/isolation invariants (no registry/library mutation, no cross-battle hook leakage, no global instance growth, no Divine exposure through Human builds), and bounded performance sanity (stress battles terminate; no growth across 30 battles). 855 tests, 82.69% coverage.
- Documented finding M16-AUD-002: targetless ULTIMATE actions fall back to the actor per the docs/04 §4.3 effect normalization (legacy defaulted to the opponent); not silently fixed — a target-defaulting rule is a
- Deliverables: MILESTONE16_DECISION_RESOLUTION.md + MILESTONE16_FINAL_COMBAT_SPECIFICATION_REPORT.md. ### Added
- Full 50-document specification set (Developer Constitution through AI Development Roadmap).
- Complete AI ecosystem documentation (Combat Intelligence through Multi-Agent Architecture).
- Repository structure: README, INDEX, appendices, AI context files.
- **Milestone 0 — Project Foundation (v0.1.0).** Layered `dio` package implementing the docs/03 architecture: - Registry framework with automatic discovery, duplicate protection, reference/circular-dependency validation, and the ten documented registries (docs/05, docs/17). - Typed event system with priority ordering, cancellation, and chain-depth protection (docs/18). - Canonical 30-hook catalog with the documented priority tiers (1000/500/300/100/0) and a HookDispatcher over the event bus (docs/06). - Trigger framework: conditions, validation rules, per-battle cooldowns (docs/06 §7-8, §12). - Calculation-layer framework: Calculator interface, Modifier model, modifier pipeline (Base→Flat→Percentage→Multiplier→Cap→Final), formula registry (docs/04 §13-14, docs/08). - Configuration (pydantic-settings, `DIO_*` env + `.env`, nested groups, feature flags), structured logging with correlation ids, error hierarchy per docs/04 §20, and a DI container. - Data layer: SQLAlchemy 2 models for every docs/13 entity (15 tables), generic typed repositories with immutable-model enforcement, Alembic migration 0001 (verified up/down, zero model drift). - Combat engine skeleton: battle phase machine, combat state, action object, action resolver contract (execution pipeline is Stage 1). - Discord bootstrap: DioBot, cog auto-discovery, graceful shutdown, fail-fast token check, `/ping` health cog (docs/16). - `dio` CLI (check/ db-init/ db-migrate/ db-revision/ bot-run), Makefile tooling, GitHub Actions CI (ruff + mypy + pytest on 3.11–3.13), pre-commit hooks. - Reference prototype preserved verbatim at `src/dio_engine/` (41 artifacts, 79 abilities; fuzz harness verified clean). - 128 tests, 85.7% coverage; new tracking docs (PROJECT_ROADMAP.md, TASKS.md, KNOWN_ISSUES.md, DECISIONS.md); AI_MEMORY.md session 2. ### Changed
- **v0.2.0 — Milestone 1: Core Combat Engine.** Added the combat engine on top of the foundation (no foundation rewrite): - Battle lifecycle, CombatState, TurnEngine, ActionResolver (11-step docs/06 §10 pipeline), EffectResolver with the documented damage and curse pipelines, status runtime (docs/07), resources, cooldowns, temporary modifiers, seeded RNG, death/victory/draw, battle records, recovery architecture, BattleRecorder persistence, DB-backed session store. - Documented calculators registered: stats.ratio_allocation, damage.final, healing.final, curse.buildup, defense.effective_ resistance, stat.modified, resource.update, cooldown.final. - Conflict resolutions (developer-approved 2026-08-10): BAL-001 spec curse formula (threshold configurable, default 200), default Human stats activated, dotted IDs confirmed, canonical hooks confirmed. - 135 new tests (264 total, 87.0% coverage); engine version 0.2.0.
- Foundation API evolution (documented): CombatState gained the full canonical state; ActionResolver executes actions (previously raised UnimplementedFeatureError); effect model added to core. ### Added (v0.3.0 — Milestone 3: Core Game Content & Registration System)
- Content architecture (`dio.content`): AbilityDefinition, ArtifactDefinition, AffinityDefinition, CurseDefinition, ConsumableDefinition, UltimateDefinition, SummonDefinition, HumanClassDefinition; Race/HumanBeing/DivineBeing configuration with the authoritative AI_CONTEXT slot limits; build rules per docs/13.
- Registry: CLASS and SUMMON content types (12 registries); production content homes wired into discovery; canonical ID prefix validation.
- ContentLibrary: typed retrieval, cross-registry dependency validation, being/loadout validation, content-version recording; DI-wired.
- Combat Engine integration: ParticipantLoadout + loadout ownership, artifact/affinity hook binding, affinity stat modifiers, curse activation content effects, consumable use limits, summon identity, being→participant adapter. Approved engine rules untouched.
- 57 new tests (345 total, 88.2% coverage). ### Changed
- Engine version 0.3.0. Registry now rejects id prefixes that do not match the registry's content type (enforcement). ### Added (v0.4.0 — Milestone 4: Production Game Content & Character Build System)
- Production content (96 items) from the existing project specification: 16 affinities, 16 curses (7 with translated mechanics), 16 ultimates (identity-only), 12 consumables, 11 supporting statuses, 25 Human Classes (tiers, roll weights, Rune Master bonuses, stat modifiers, canonical-hook mechanics).
- Class Roll System (TIER_WEIGHTS 44.85/44.95/10/0.1) via the centralized RNG.
- Character/build persistence: CharacterService over the existing schema; migration 0002 (character affinity, docs/13 §6); affinity-ultimate gating; cumulative-stats/ratio validation.
- 42 new tests (387 total, 87.8% coverage). ### Changed
- Engine version 0.4.0. Engine changes were minimal content-integration additions (status data payloads, class stat modifiers at preparation, class hook binding, BattleRNG.choices/sample). ### Added (v0.4.1 — Milestone 5: Combat Content Integration & Spec Resolution)
- Full audit/design milestone: content matrix, curse/ultimate/harmony/ consumable/status/artifact/summon/nullification audits, Decay conflict investigation, engine extension proposal, implementation boundary classification.
- Audit-justified infrastructure: explicit approved combat stage order (dio.engine.lifecycle), CombatContext standard content context (dio.engine.context), summon-ultimate restriction guard (documented rule), declarative legacy-hook map (48 hooks, dio.content.legacy.hook_map). 11 new tests (398 total, 88%).
- MILESTONE5_COMBAT_INTEGRATION_REPORT.md with ### Changed
- Engine version 0.4.1. No approved combat rule or production content changed; blocked mechanics remain blocked and documented. ### Added (v0.4.2 — Milestone 5 corrections: Canonical Decisions & Harmony Design)
- Canonical.): Decay (60% healing->damage canonical), Nullification absolute negation (implemented in the EffectResolver; bypass only when explicit), Summons never trigger artifact passives (verified), Rune Master confirmed, connected Affinity/Curse/Ultimate/Harmony model, Human/Divine content separation across the whole content ecosystem (current set classified Divine), 25-50 total-action pacing.
- Harmony system implemented (dio/engine/harmony.py): four approved generation sources wired into battle events and the resolver, economy designed around the pacing model, ultimate Harmony requirements (low 40/ moderate 60/ high 80/ very_high 100) and spend-reset-to-0.
- Human/Divine separation: production content classified; filtering + validation; tests.
- 15 new canonical-rule tests (416 total, 88% coverage). ### Changed
- Engine version 0.4.2. Engine changes were the audit-justified corrections above; no approved combat rule or balance value changed outside the user-authorized Harmony economy design. ### Fixed
- Foundation-era defect: `apply_default_human_stats` behavior was pending approval — now approved and implemented per AI_CONTEXT.
- Hook integration gap closed: before_damage/after_damage/before_heal/ after_heal now emit at the documented moments.
- **Final approved rules implemented (2026-08-10, DECISIONS ..):** crit multiplier caps per race (Human 4x/3x, Divine 3x/3x) with explicit overrides; summon source identity + canonical summon events; Basic Attack = Power x RNG(0.85-1.25) via the seeded battle RNG; Rest = +3 Energy, +20% lost HP, +20% DR until the player's next turn; curse meter resets on activation with no default overflow (per-effect override available). BattleConfig.basic_attack/rest_recovery placeholders removed; 25 new tests (288 total, 87.4% coverage). ---

- **Approved change:** the Light Rune's restriction is now represented as a RUNE-LEVEL eligibility rule declared on the Rune itself, rather than hard-coding Light Rune eligibility as a generic class restriction. `RuneDefinition` gains `eligible_classes` (empty = every Human class permitted by the Rune's own rule) and `eligibility_reason`; the loader enforces the Rune's rule IN ADDITION to the existing raid-exclusive layer ( Human classes, RuneMaster, and all class mechanics are fully preserved; no class content changed.
- **Approved rule ( Divine beings cannot (Human-only Rune, `content_race`). The raid-exclusive/RuneMaster access layer remains enforced on top, so the effective equipping class set is unchanged (Human classes with raid access). The other 15 Runes are untouched (no Rune-level rule declared).
- Engine: `eligible_classes`/`eligibility_reason` on `RuneDefinition` + `is_eligible(being)`; `_rune` builder passes the rule through (always-serialized `eligible_classes` metadata key); loader validation (Rune-rule check + declared-class-id check) in `validate_being`.
- Tests: `tests/unit/test_m24_5_light_rune_eligibility.py` (14 tests: Rune-level declaration + metadata, every Human class passes the rule, Divine rejected, raid layer on top, RuneMaster preserved, named-list enforcement, unknown-class diagnostics, other 15 Runes unchanged, serialization round-trip, registry integrity); M23 audit allowed-list updated (documented).
- 1032 tests (1018 + 14)/ 83.36% coverage/ mypy strict clean (116 files)/ Ruff clean (212 files)/ build `andy_engine-0.15.1`/ registry 269 (unchanged)/ `dio ready` exit 0/ fuzz 0 errors/ secret scan 0 hits. No deployment.

### Added
- **Human Rune system** (the approved 16 Runes; Artifact subtype, no parallel system): `RuneDefinition` in dio.content.base with `bound_affinity`/`bound_curse`/Bonded-Unbonded tiers; 16 Runes registered in the artifact registry (40 -> 56) with machine-verified Affinity -> Curse bindings (AFFINITY_CURSES), `content_race="human"` and `raid_exclusive=True`.
- **`raid_exclusive` artifact capability** + RuneMaster access: `ArtifactDefinition.raid_exclusive`; the production RuneMaster now carries `grants_raid_artifact_access=True` (the production port of the legacy flag); loadout validation rejects raid-exclusive artifacts for unauthorized classes and validates Rune bindings.
- **Rune effects (v1 values, documented revisable):** common lever = bound-curse buildup amplification (Bonded x1.30/ Unbonded x1.15, before resistance); Verdant +2/+1 seeds per activation with INFINITE accumulation ( technical bound only; the parasite "seeds" data key now accumulates additively); Sanguine Anemia 28/65 +5% growth (Bonded)/ 24/55 +3% (Unbonded) with INFINITE battle-local growth (serialized in the new per-combatant `rune_state`); Light Starvation Harmony drain 25/22 (no new Harmony source); Entropic curse-activation duplication (Bonded any curse/ Unbonded bound curse) with NO gameplay chain cap (the existing engine event-depth safety boundary is preserved); Boreal +1 Permafroze charge (Bonded).
- **Acquisition (approved model):** `RuneDropService` — solo 10%, group 1%, exactly ONE group roll, battle-id ledger duplicate protection, victory-eligibility interface; group distribution is recorded as PENDING_GROUP_DISTRIBUTION (open decision).
- **PvP reward interface:** `PvPRewardEvaluator` (legitimate-victory + anti-repeat + config-gated minimum rating); rating NEVER grants combat power. Reward settings (an internal setting/GROUP, an internal setting) in the config.
- **Wildcard infrastructure:** PvE/Boss-only reward items; PvP use rejected; effects not finalized (future milestone).
- 41 new tests (tests/unit/test_m24_runes.py): registry/access/bonding/ effects/acquisition/PvP/wildcards/isolation/serialization/replay. 1018 tests, 83.33% coverage. ### Changed
- `CombatantState.rune_state` (serialized battle-local Rune state); `StatusManager` count-stacking additively merges the "seeds" data key; the parasite seed count now scales with accumulated seeds (aligned with the documented per-seed intent; identity without a Rune). ### Not done
- No Divine Runes; no unauthorized Human systems; no economy/ progression; no deployment. Rune numbers are v1 and may be re-audited after the remaining Human systems are integrated.

### Added (audit artifacts only — no production change)
- Audit of all 16 proposed Runes (machine-verified Affinity->Curse pairs, curse mechanics, artifact architecture, RuneMaster, Harmony economy); deep audits for Entropic, Verdant, Sanguine, Light.
- Decision package with ..reserved + the architecture decision (raid_exclusive/ acquisition).
- 17 audit tests (tests/unit/test_m23_rune_audit.py); 977 tests, 83.14% coverage; version stays 0.14.1.
- Documented gaps: Appendix K absent from the repository; full proposal texts not supplied; Ruin/Physical premises mismatched. ### Not done
- NO Rune implemented, registered, activated, or made canonical; no production behavior changed; no engine capability added; no deployment.

### Added
- Final transferable release archive `release/ANDY_Replit_Release_v0.14.1.zip` (1,202,476 bytes; SHA-256 3ec3bec4bb98db8822368db4d1956521d41b310ccddeadd0e287f395b29eaf4e) with root directory `ANDY_Replit_Release_v0.14.1/` — 364 files, 53 directories; no secrets, caches, local databases, or build artifacts.
- Handoff set: `ANDY_RELEASE_MANIFEST.md`, `ANDY_FILE_INVENTORY.md`, `ANDY_REPLIT_QUICK_START.md`, `ANDY_RELEASE_SHA256.txt`, `MILESTONE22_6_USER_FILE_HANDOFF.md`.
- Verified from the staged copy and from an independent extraction of the ZIP: 960 tests, 83.14% coverage, mypy/Ruff clean, migrations at head, registry 253, `dio ready` exit 0, prototype fuzz 0 errors, secret scan 0 hits. ### Not done
- No code changes (version stays 0.14.1); no gameplay changes; no deployment; no Discord connection; no production database; no credentials added. The archive is the handoff — Replit import, Secrets, PostgreSQL, migrations, bot start and live validation are operator actions.

### Changed
- **Project identity renamed DIO → ANDY.** User-facing branding now says ANDY (README, current-state docs, CLI descriptions, the `dio check` app line via `settings.app_name` default `andy`, Dockerfile comments, Replit packaging, the M22 Replit handoff, and the ops docs).
- Python distribution metadata renamed `dio-engine` → `andy-engine` (pyproject name/description/authors/keywords). Version remains **0.14.1** (no version bump for a rename).
- Replit packaging refreshed for ANDY (`.replit`, `replit.nix`, REPLIT_DEPLOYMENT.md, REPLIT_ENVIRONMENT.md, REPLIT_PRODUCTION_CHECKLIST.md); the run command stays `python -m dio bot-run` (retained internal entry point). ### Retained (documented compatibility contracts)
- Internal Python package/import path `dio`; CLI `python -m dio`; `DIO_*` environment variables (`an internal setting`, `an internal setting`, `an internal setting`, …); database schema/tables; migration identifiers (head a commit); serialized state keys; replay format; registry/content IDs (253 items unchanged); Discord command identifiers; error codes. ### Not done
- No gameplay/content/balance changes; no registry change; no database migration; no production deployment; no Discord connection; no credentials added; no new systems.

- **Deployment verdict C — DEPLOYMENT BLOCKED** (no code changes; version stays 0.14.1). The workspace is not a Replit repl; no an internal setting/ an internal setting/ an internal setting; no PostgreSQL/git. No live deployment occurred (M22 safety rule).
- Added the Replit deployment package (configuration artifacts): `.replit`, `replit.nix`, REPLIT_DEPLOYMENT.md, REPLIT_ENVIRONMENT.md, REPLIT_PRODUCTION_CHECKLIST.md, plus MILESTONE22_REPLIT_PRODUCTION_HANDOFF.md and MILESTONE22_REPLIT_PRODUCTION_VALIDATION_REPORT.md.
- Added 12 production-boundary tests (tests/unit/test_m22_production_validation.py). 939 tests, 83.09%.

- **Deployment verdict C — DEPLOYMENT BLOCKED** (no code changes; version stays 0.14.1). All environment-free gates passed (927 tests, 83.09%, ready exit 0/1-paths, bot dry-run, fuzz, secret scan, production guards verified at runtime, session-store persistence round-trip). The deployment chain stopped at the infrastructure gate: no Docker/PostgreSQL/host, no credentials, no git identity. See MILESTONE21_PRODUCTION_DEPLOYMENT_REPORT.md.

### Fixed
- **M20-AUD-001:** Soul Manipulation's ON_DEATH restore now applies the authoritative legacy power formula (`base_power += (max_power − power) × 0.10`). The ported implementation computed `power − power` (always 0) — a genuine incorrect implementation of an already-authoritative specification (the power-restore component was a silent no-op). No balance change: the effect now matches the legacy exactly. Regression-tested. ### Added
- Final audit layer (14 tests in tests/unit/test_m20_final_integration.py): combat lifecycle edges (turn-start DoT death, TURN_END kill, draw path, summon owner death, ended-battle rejection), serialization round-trip of every stateful field, same-seed/different-seed determinism, 30-battle isolation, security boundaries, unauthorized-system guards, and the Soul Manipulation regression. 927 tests, 83.09% coverage.

### Added
- **Human Class mechanics completed** (the authoritative Human content set — 25 classes from the legacy `classes/` package): - Data-driven (already present, verified): Knight (+20% HP, +10% DR), Assassin (−30% HP, +1 crit mult, +20% dodge), Spirit King (full curse immunity), Rune Master (slot bonuses). - Hook mechanics completed: Swordsman (−20% curse inflict + 20% attack re-trigger, chain ≤ 5), Spearman (DR piercing), Archer (1-turn attack delay + passive-attack immunity), Martial Artist (back-to-back +50%, the legacy documented approximation), Mage (+20% curse inflict), Summoner (20%-stats summon every 5th owner turn), Enchanter (ally stat transfer on ally turns), Healer (50% heal share to allies), Spellblade (+25% attack AND curse damage), Battle Mage (affinity-curse trigger — BLOCKED component: no Human affinity content exists), Paladin (+50% healing received), Fallen Knight (+75% attack damage, +50% curse taken), Necromancer (battle-start undead; summon control stays blocked), Dragon Tamer (battle-start dragon, 10 Purgatory/turn, dragon on curse trigger), Time Weaver (opponent attack delay), Blood Lord (uncapped lifesteal), World Tree Guardian (stat-debuff immunity).
- **Minimal engine capabilities (all item-driven, battle-local, deterministic, serialized):** - `uncapped` healing flag (, Blood Lord) — skips the max-HP clamp and the Decay conversion (legacy direct health add). - Deferred-attack machinery (, Archer/ Time Weaver): `CombatantState.pending_attacks` (serialized) + resolution at the combatant's next turn start; `Battle.should_defer_attack`/ `basic_attack_effects`/ `_resolve_pending_attacks`. - `CombatState.last_manual_attacker` (, Martial Artist; recorded after the attack's effects resolve — legacy order). - Damage-origin markers: ABILITY damage is stamped `is_ability`, ULTIMATE/ARTIFACT/ITEM damage `is_manual` (legacy deal_damage defaults), so Archer's passive immunity distinguishes action damage from hook/status/summon damage.
- 26 new tests (tests/unit/test_m19_human_content.py): every class mechanic, registry/roll/validation, determinism, battle isolation, Human/Divine separation, security. 913 tests, 83.02% coverage. ### Fixed
- None (no pre-existing defect was found; the M19 audit confirmed the class-registration infrastructure).

### Added
- **Decision gate resolved (..):** the user approved five mechanics and rejected/kept four: - **Voidpiercer's Edge crit cap:** new `crit_multiplier_cap` per-combatant override channel; the legacy 5.0 raise is applied. Silent Reckoning stays at the 3x (the legacy clamp itself capped its 6.0 spike at the default 3.0 — 3x is the faithful legacy outcome). - **Targetless ultimates (, M16-AUD-002 resolved):** a targetless ULTIMATE action defaults to the actor's primary opponent (legacy parity), resolved before validation so validation/hooks/context/effects/replay agree. Self-targeting ultimates unaffected; actor fallback remains for other action types. - **Deathless Vigil resurrection:** true battle-local resurrection via ON_DEATH — full HP, full power (modifiers cleared), death fields cleared, once per battle; −30% max HP at the next owner turn start. Replaces the M6-era BEFORE_DEATH translation. - **Morphogenesis resistance shuffle:** new per-curse resistance channel `curse_resistance.<id>`; the user-confirmed factor-preserving mapping r = clamp(1 − 1/(1+U), 0, 1) with U ~ uniform(−30, 30) per landed manual attack (legacy on_manual_attack gate). Documented divergence: U < −1 clamps to immunity (legacy drained the curse meter). - **Split Personality:** user-defined pool = the Divine registry (active, unowned, cap 5, seeded pick); battle-local grant at the legacy −5 energy-cost cut. No levels (UD-01 rejected), no auto-use (UD-02 rejected).
- 19 new tests (tests/unit/test_m18_decision_resolution.py) covering every approved mechanic, the rejected-decision locks, and the inventory/matrix state. 887 tests, 83.01% coverage. ### Fixed
- None (no pre-existing defect was found in the M18 audit).

### Added
- **Decision gate:** the M17 brief granted no approvals of M15-UD-01..08 or M16-AUD-002; all remain (content unchanged); M15-UD-09 (deferrals) and M15-UD-10 (Human unauthorized) preserved.
- **Final audit layer** (13 tests in tests/unit/test_m17_final_audit.py): decision-gate locks (no restore/rewind/rollback surface, no extra-turn machinery, no ability levels/XP/rarity on ANY registered ability, no Human content in the prohibited registries, mandom/endlessly unregistered, Ability Cache inert+blocked), the M16-AUD-002 pinned behavior, a 5-seed multi-seed determinism sweep, and bounded stress re-verification. 868 tests, coverage 82.99%. ### Fixed
- **M17-AUD-001:** a TURN_END hook's effects could kill the last opponent (e.g. Mirror Protocol's turn-end strike), ending the battle inside the hook resolution (resolve_effect -> _check_deaths -> check_victory -> _end_battle); `TurnEngine.end_turn` then crashed with BTL-608 ("cannot end a turn: no active turn") because the end-turn phases ran on the already-ended battle. `end_turn` now returns cleanly when the battle ended during TURN_END hooks; the victory/BATTLE_ENDED/session-finalize recording was already done by `_end_battle`. Regression-tested (minimal reproduction + the mixed-content path that originally exposed it).

### Added
- **`mode="set"` stat effect:** STAT_MODIFY metadata flag that turns the effect into an absolute set-to-value -- the resolver computes the compensating flat delta (target - current effective, divided by the active percentage x multiplier factor) so the channel's effective value equals the requested value; `hp` sets the field directly, clamped to the effective max_hp. This is a stat modification ONLY -- no snapshots, history, or rollback (the restore-engine prohibition is untouched).
- **Cooldown validation gate:** the action-validation cooldown check no longer short-circuits on a declared `meta.cooldown`; a COOLDOWN_SET effect may place ANY content on cooldown (content without cooldown entries remains always-ready, so existing behavior is unchanged).
- **Echoed Sanctuary:** the last restore-family artifact -- stats capture/restore of the eight legacy Stats fields ONLY (+10% while captured, 20-turn auto-revert, manual revert on second activation, 5-turn cooldown). Legacy scope verified from the source: `copy.deepcopy(owner.stats)` -- no statuses/curses/cooldowns/roster/ harmony/energy/summons.
- **Sundering Verdict:** battle-start disable of one random opponent ability per opponent (legacy `disabled=True` translated as a battle-long COOLDOWN_SET -- activation blocked, stays equipped, hooks keep firing, matching the legacy dispatch which never filters disabled) + one random artifact destroyed (ROSTER_MODIFY) + Crush 50; dodge-gated activation removes a random opponent ability + Crush 50.
- **Fortunes Undertow completed:** per-turn percentage modifiers on BOTH power and max_power (multiplicative compounding reproduces the legacy compounding exactly); the Moper catch scales all four pool stats.
- **Prismatic Riposte completed:** full legacy dodge pipeline -- pools x1.25, seeded uniform 0.1-0.9 re-split, then ALL eight stats x1.25; the legacy dodge-cap raises are moot under the production cap (1.0 = always dodge).
- **Ultimatum reclassified PORTED:** the legacy temporary passive-disable (disabled=True on the target's passives) is behaviorally INERT in the legacy engine -- `disabled` is consulted only by `can_activate` (passives are never activated) and the legacy hook dispatch never filters disabled abilities -- so the production port already captures the complete legacy behavior.
- 28 new tests (tests/unit/test_m15_final_content.py): the four completions, the set-mode capability, race isolation, battle-local state, serialization, determinism, nullification scope, summon handling, deferral/block regressions, and the repository-computed completion matrix. ### Fixed
- None (the M15 audit found no pre-existing engine bugs; the two capability changes are documented extensions,/).

### Added
- Battle-local roster operations: EffectType.ROSTER_MODIFY (grant/remove/remove_all of abilities/artifacts; seeded random selection; hook bind/unbind; ABILITY_GRANTED/ ABILITY_REMOVED/ ARTIFACT_DESTROYED events emitted before unbinding; battle-local only).
- Registry-random ability grants: Battle.random_ability.
- Unblocked abilities (4 new + 5 completed): system_breach, soft_and_wet, dreaded_return (summon resurrection), jackpot (partial: no nested auto-use), mutation (refund), cyberpsychosis (scramble), ultimatum (cost halving), colossus_unbound (wipe).
- Unblocked artifacts (5): aeonic_flux, reapers_gambit (pool flips via the max-stat channels), sovereigns_genesis, release_schedule (partial: no level growth), communal_reprieve (awaken heal).
- All 16 curses implemented: parasite (drains + permanent heal redirect), immunity (copied resistances + inflictor-scoped countdown), cataclysm (artifact destruction), plague (Antibody strikes on the inflictor's turn start).
- 26 new tests (tests/unit/test_m14_content_completion.py): roster capability, serialization, per-item behavior, complex replay + isolation, deferral regressions. ### Fixed
- Cross-battle state leak: the Plague antibody id counter lived on the shared registry-level curse instance; ids are now derived per battle.
- Roster removal now emits the removal event BEFORE unbinding the removed content's hooks (Mutation's refund and Sovereign's Genesis' replacement must observe it).

### Added
- Max-stat modifier channels: `stat.max_hp`/ `stat.max_power` modifier-backed; effective max-HP/Power used by healing caps, percent-max damage, Rest, hp_pct, and the max-HP ultimates; current pools clamp to new maxima (legacy semantics).
- Ability-cost modifier channel: `ability_cost.<content_id>` flat/percentage modifiers applied in the resolver (validation and spending agree); STAT_MODIFY accepts a `target_calculation` override.
- Snapshot round-trip includes the modifier bank.
- Abilities completed (10): minazuki, subliminal_light, conqueror, desperation, devils_rhythm, dark_determination, heros_valor, aegis_protocol (partial -> ported), adaptation (blocked -> ported), overdose (blocked -> ported).
- Artifacts completed (3): inertia_reversal, surging_current, radiant_undertow (partial -> ported).
- 22 new tests (tests/unit/test_m13_completions.py): engine channels, per-item behavior, Mandom/Endlessly deferral regressions, canonical rules intact. ### Fixed
- `CombatState.from_dict` restored no modifiers (serialization gap; ).
- Percentage-modifier composition clarified as multiplicative -- the completions match the legacy compounding exactly.

### Added
- Production configuration guards: SQLite + debug rejected in production; log-level validation; Discord token `repr=False` (secret redaction).
- Bot startup hardening: config validation, database reachability (APP-104), and schema presence (APP-105) are verified before any Discord connection; migrations stay an explicit step.
- Graceful shutdown: close closes Discord then disposes the database pool; never raises.
- `dio ready` readiness probe: config + database + registry, exit 0 = READY.
- Containerization: Dockerfile (slim, non-root, no baked secrets, `dio ready` healthcheck).dockerignore, docker-compose.yml (PostgreSQL + optional bot profile).
- Operations docs: DEPLOYMENT.md, OPERATIONS.md, SECURITY.md, PRODUCTION_CHECKLIST.md.
- 24 new tests (config validation/redaction/production rules, startup failure modes, graceful shutdown, persistence-failure safety, recovery concurrency, CLI readiness, smoke bootstrap). ### Changed
- Settings.app_version now defaults to the package version (previously stale 0.1.0).
- Bot startup with a missing token now fails via the authoritative configuration validation (CFG-100).

### Added
- Application services (Discord-independent): BattleService (battle creation, engine-authoritative submission, turn-loop driving, finalization/persistence), ChallengeService (ephemeral, concurrency- safe accept/decline/expiry), ContentService (registry-backed browsing), RecoveryService (safe session recovery).
- Discord layer: ProfileCog (/profile,/character,/build), ContentCog (registry-driven browsing commands), PvpCog (/challenge,/recover,/recover_end); ChallengeView (Accept/Decline); BattleView (Attack/ Rest/ Ultimate buttons + Ability/Artifact/Consumable/Target selects generated from production state); battle embeds; DiscordBattleSession Manager (battle identity, participant ownership, per-battle action locks, stale/duplicate/ended rejection).
- Container wiring for all new services (bootstrap.py).
- No new database migration: sessions use the battles table status; players link via discord_id. ### Changed
- BattleService.submit advances the turn after a successful action; rejected actions never advance. ### Fixed
- Ultimate button resolves the actor's ultimate from the loadout.

### Fixed
- Action-restriction stacking now additive (legacy Seal overflow: turns_left + 2 per activation; ).
- Seal curse preserves its blocked rows on re-activation while the target is still sealed (legacy overflow semantics; ).
- Stale "placeholder" comments about curse meter consumption corrected (the mechanism was already implemented). ### Added
- 21 audit tests (tests/unit/test_m10_audit.py): lifecycle edges (rejected ultimate harmony, multi-effect AFTER_ULTIMATE, ultimate exemption from restrictions), Decay interaction edges (lifesteal, ultimate heal), summon audit, strong-attack matrix, reflection one-chain + reflection-vs-nullification, absorb/nullify distinctness, complex combined-battle deterministic replay + battle isolation, Curse Vial with identity-only curses, hook-id uniqueness, bounded heavy-battle performance sanity.
- MILESTONE10_COMBAT_INTEGRATION_HARDENING_REPORT.md (readiness: READY FOR DISCORD INTEGRATION).

### Added
- All 16 Ultimate V1 designs implemented exactly as user-approved: Ember Cremation, Boreal Absolute Zero, Voltaic Thunderstorm, Abyssal Abyssal Decree, Tidal Tsunami, Light Radiant Dawn, Arcane Arcane Binding, Ruin Ruin, Verdant Verdant Harvest, Noxious Pandemic, Physical Iron Body, Entropic Entropic Cascade, Gravitic Singularity, Celestial Celestial Chorus, Spectral Ghost Parade, Sanguine Blood Moon (`src/dio/content/ultimates/__init__.py`; `spec_status=canonical_v1`).
- Approved Harmony requirements as content costs (60/80/100 per the approved table); activation spends + resets Harmony.
- 4 new ultimate-support statuses (status.ultimate_absolute_zero, status.ultimate_iron_body, status.ultimate_entropic_cascade, status.ultimate_blood_moon); status count 42 -> 46.
- 40 new tests (tests/unit/test_m9_ultimates.py): per-ultimate behavior, activation requirement, harmony consumption, affinity gating, target validation, damage/healing, status effects, curse interaction, Nullification, DR, Absorb, Reflection, dodge, summon interaction, deterministic replay, battle isolation.
- MILESTONE9_ULTIMATE_IMPLEMENTATION_REPORT.md.
- MILESTONE8_ULTIMATE_PROPOSALS.md status updated to CANONICAL — VERSION 1 (approved 2026-08-10). ### Changed
- test expectations updated from the pre-approval identity-only state to the approved canonical state (documented expectation change, not a weakened assertion). ### Fixed
- (none — no regressions; 638 baseline tests unchanged and passing)

### Added
-.recorded and regression-tested: - : ported Ability/Artifact content remains DIVINE. - : Mandom stays DEFERRED (no state-restore engine). - : Endlessly stays DEFERRED (no extra-turn engine). - : Decay uses the production damage pipeline (CONFIRMED). - : Ultimate effects may be drafted as proposals only.
- Engine capabilities (documented extensions): - Status-attached hooks + status application effects (StatusDefinition.declare_hooks/ apply_effects; ). - Ability-cost discount statuses (StatusDefinition.ability_cost_discount; ) — Overclock Chip. - Last-action summary + last-curse-inflicted tracking on CombatantState — Spectre curse, Echo Charm.
- Consumables completed: Iron Ward, Ghost Draft, Vampiric Thorn, Overclock Chip, Stat Tonic, Cleansing Draft, Focus Shard, Echo Charm (CONSUMABLE-001/003 resolved).
- Curses completed: Permafroze (status-attached doubling), Echo (engine-integrated spread), Spectre (last-action replay) — 12 of 16.
- Abilities: King Crimson's same-ability-twice lock completed via status.kc_lock (56 PORTED).
- Artifacts: Hexward Sanctum (per-turn meter decay) and Scarecrow's Toll (full meter cleanse) completed (23 PORTED).
- New shared statuses: status.curse_permafroze, status.curse_echo (rebuilt), status.kc_lock (42 total).
- MILESTONE8_ULTIMATE_PROPOSALS.md — 16 Ultimate design proposals, all PENDING USER APPROVAL.
- 37 new tests (tests/unit/test_m8_decisions.py, test_m8_content.py). ### Changed
- StatusManager.apply returns (active, created); status-attached hooks and application effects bind only for new instances (no duplicate firing on stacking).
- Action resolver applies and consumes ability-cost-discount statuses. ### Fixed
- (none — no regressions; 600 baseline tests unchanged and passing)

### Added
- **Decay conversion hook**: the canonical 60% healing->damage rule is implemented end-to-end (curse.decay activation -> status.curse_decay -> engine conversion in the healing resolver, data-driven via StatusDefinition.healing_conversion); CURSE-MEC-002 and DECAY-HOOK-001 resolved.
- **Action restrictions**: EffectType.ACTION_RESTRICT + the action_restricted event; enforced by the resolver (docs/18 §12); Seal curse, Silent Reckoning stagger and Scarecrow's Toll punish now functional; Right of Conquest gains rage on restriction and shatters restrictions at rage>=70.
- **One action per turn**: the resolver enforces the documented turn structure (docs/18 §9); rejected actions do not consume the turn.
- **Strong-attack derivation**: a landed non-critical hit above the attacker's power is "strong" (DOC-002 resolved); Gold Experience Requiem's strong defense triggers correctly.
- **Tests**: 30 new (600 total; 83.49% coverage) -- Decay conversion, restrictions, one-action-per-turn, exactly-once AFTER_* events, dodge integrity, strong derivation, nullify/absorb/reflect distinction, battle isolation, deterministic replay with passives, pacing measurements. ### Changed
- Version 0.5.0 -> 0.6.0; status count 39 -> 40 (status.curse_decay); implemented curses 7 -> 9 (Decay, Seal).
- Tests updated to model real turns (one action per turn) -- no assertions weakened. ### Documentation
- DECISIONS .., KNOWN_ISSUES (CURSE-MEC-001/002, DECAY-HOOK-001, DOC-002, PORT-DEV-001, M7-GAP-001, M7-USER-001), TASKS/ROADMAP/ AI_MEMORY updated, MILESTONE7_COMBAT_COMPLETION_REPORT.md.

### Added
- **adapter** (`dio/content/legacy/`): the translation layer from the preserved prototype to the production engine -- legacy name maps (16 curses, 28 statuses), effect-building helpers (damage/healing/ curse/status/stat/energy/summon), deterministic RNG passthroughs, and the definitive 79+41 content inventory (`inventory.py`) with per-item PORTED/ PARTIAL/ BLOCKED status and deviation notes.
- **Ported content**: 70 of 79 abilities and 34 of 41 artifacts registered as DIVINE content (55+21 PORTED, 15+13 PARTIAL with documented deviations, 9+7 BLOCKED with exact reasons); 28 supporting legacy statuses (shared infrastructure).
- **Engine extensions (consumer-justified)**: 9 new canonical events/hooks (dodge, curse buildup/activation/before, status apply before, stat modify before/after, energy gain/spent/lost/before); AFTER_ATTACK/ABILITY/ULTIMATE/CONSUMABLE/REST events now emitted with outcome data; `ActionType.ARTIFACT` (active artifact activation at the Artifact stage); dynamic costs (`effective_costs`) + one-use `_temp_discount`; negative curse buildup = meter reduction; owner- scoped content instances; ability hook binding; kill-by-crit/ last- hit-almighty attribution; `prevent_dodge` before-attack flag.
- **Tests**: 154 new (570 total; 83.28% coverage) covering the adapter, engine extensions, every ported ability/artifact, deep artifact branches, and cross-system rules (summon/artifact restrictions, nullification, race separation, determinism, harmony/energy correctness). ### Changed
- Version 0.4.2 -> 0.5.0; registry counts: 70 abilities, 34 artifacts, 39 statuses (production content, DIVINE); inventory-driven.
- Resource-event decision (RES-EVT-001) documented exception for energy events. ### Documentation
- DECISIONS .., KNOWN_ISSUES HOK-EXT-001/PORT-DEV-001/ DECAY-HOOK-001, TASKS/ROADMAP/AI_MEMORY updated, MILESTONE6_HOK001_CONTENT_PORT_REPORT.md.

### Added
- `docs/01` – `docs/09`: Core engine architecture, combat design, registry, hooks, status effects, formulas, balance philosophy.
- `docs/10` – `docs/12`: Combat Intelligence, Theory Engine, AI Design Philosophy.
- `docs/13` – `docs/22`: Database schema, testing, security, command architecture, content pipeline, event system, AI training pipeline, progression, analytics, matchmaking.
- `docs/23` – `docs/48`: Full AI ecosystem — decision engine, counterplay discovery, knowledge network, self-play, reinforcement learning, personality modeling, coaching, live assistance, memory management, explainability, balance analysis, content generation, tournament analysis, anti-cheat, world simulation, autonomous research, multi-agent collaboration, communication, long-term evolution, dataset governance, performance optimization, failure handling, developer administration, user assistant, competitive ranking, future expansion.
- `docs/49`: Complete AI Architecture Integration Document.
- `docs/50`: AI Development Roadmap & Implementation Priority. ---

- This changelog tracks **documentation and specification** versions, not necessarily deployed code versions.
- Code implementation versions should be tracked separately once implementation begins, per `docs/50_AI_Development_Roadmap_Implementation_Priority.md`, Section 11.
- Major AI behavior changes must include migration notes and rollback procedures per `docs/41_AI_Long_Term_Evolution_Self_Improvement_Framework.md`. ---

- **MINOR bump:** Lucifer, Mammon, Asmodeus, Nemesis, Beelzebub, Satan, Belphegor — Divine superbosses, each with a unique Divine Law implemented entirely through the existing hook/event system (no new combat subsystems; isolated to Sin ids).
- Laws: Pride (30% Power cap + Excess→All-Stats, Ascension, Challenge Marks, Pride Memory, Non Serviam), Greed (Buff/Artifact/Consumable theft, Greedy Tax, Debt Collections, Mine), Lust (redirection to Lust Slaves, spawn cap + Temptation, Obsession, Desire Memory, Ecstasy), Envy (0.5x cumulative per action, Buff Mirror, Summon Prohibition, Adaptation, Mirror of the Damned), Gluttony (Curse/ Damage/Summon/Consumable consumption, Hunger→Feast, Abyssal Feast), Wrath (Rage, retaliation slots, execution below 50% HP, Wrath of God), Sloth (current-Power decay, +10pp DR per hit, Turn Drain to exhaustion, curse reflection, Apathy, Eternal Sleep).
- Engine (minimal, generic): BattleParticipant.sin_id (serialized), Sin-hook binding at battle start. No combat order/formulas/caps/ Divine content changed.
- Tests: `tests/unit/test_m30_5_sins.py` (20). **1230 total, 80.11% coverage**, mypy (138), Ruff (242), build 0.18.0, ready 0, fuzz 0, secret 0. No deployment.