chore(v20b): remove the implicit capture layer — implicit signals go, explicit engagement stays

[runyourempire]

v20b: removal of the IMPLICIT behavioral-capture layer, per the completed read-only inventory and the seven operator rulings (2026-08-17). Companion to #474 (v20a). Implicit capture goes, explicit engagement stays, scoring is byte-identical — NO PIPELINE_VERSION bump (stays at 21). New decision record: AD-031; AD-029 amended; INV-071 retired.

Rulings A1–A7 as implemented

• A1 — Dismiss / MarkIrrelevant / BriefingDismiss STAY. They still write interactions rows and feed the kept stability_detector::record_evidence loops and record_item_feedback. Only their dead affinity/anti-topic consumers were removed.
• A2(a) — SkillsDimension.top_affinities + AffinityEntry + INV-023 guard test KEPT. Only the source query in sovereign_developer_profile.rs::assemble_skills was deleted; the field is permanently empty (serialized-shape stability, same doctrine as the pinned ScoreBreakdown fields). The v20 explicit skill-gap query directly below it is untouched; all four skill-gap tests pass, including behavioral_affinities_do_not_exclude_gaps and no_explicit_engagement_means_no_gaps.
• A3 — the three affinity/anti-topic anomaly detectors removed (detect_context_drift, detect_contradictions, detect_confidence_mismatch + their tests/fixtures). detect_abnormal_volume and detect_stale_data KEPT. The AnomalyType enum variants were deliberately KEPT so historical stored anomaly rows still deserialize and render — this kept the Anomalies UI coupling trivial (no STOP needed).
• A4 — mcp-4da-server DEFERRED to v20c. No code/db changes in it. Its reads degrade gracefully on the dropped tables; note that its db.ts CREATE TABLE IF NOT EXISTS may resurrect empty topic_affinities/anti_topics tables in 4da.db — cosmetic, cleaned in v20c. (One exception, disclosed: two retired-ok HISTORY MARKERS added to mcp-4da-server/CHANGELOG.md — the check-retired-claims gate fails on the base because the 5.0.2 changelog entry quotes the retired claims as history without the gate's documented marker. Lint markers only, zero behavior.)
• A5 — bootstrap term UNTOUCHED, rows KEPT. Live probe (2026-08-17, post-chore(scoring): PIPELINE_VERSION 21 — bump-only activation of #471's word-boundary fixes #482): 713 interactions total, 296 with |signal_strength| ≥ 0.3, ALL 296 implicit ⇒ effective_feedback_count = 98 ⇒ bootstrap OFF, held off entirely by implicit rows. v20b performs NO DELETE/UPDATE on interactions, so the count — and therefore scoring — is unchanged. Open v20c question (recorded in AD-031): should bootstrap-exit require EXPLICIT signal? That is a scoring change and needs a PIPELINE_VERSION bump.
• A6 — docs: INV-071 marked RETIRED with a reason, body preserved as history (its display-gate bullet was already stale — the cited counter at ace_commands/interactions.rs:429 died in v20a; the file now ends at ~400 lines). AD-029 amended to record the 2026-08-17 partial adoption of its previously-rejected "full removal" option (implicit removed, explicit kept — the "destroys honest surfaces" objection dissolved when v20a removed those surfaces). AD-031 added following the file's format. specs/: historical-status banners added to ACE-STONE-TABLET.md and CONTEXT-ENGINE.md (matching the existing 2026-08-16 banner idiom); history sections not rewritten.
• A7 — snooze untouched. No snooze arm added, FE call not removed.

STOP facts honored

1. record_interaction / store_interaction / ace_record_interaction all survive — only the Scroll/Ignore arms and the implicit-skip block were excised inside them. ace_record_interaction stays registered everywhere (lib.rs, CommandMap, Victauri allowlist). The implicit-skip excision was scoped to the if-block inside the same if let Ok(conn) as the KEPT stability loops.
2. ACE DB = main DB honored throughout (Phase 105 runs on the shared 4da.db).
3. Scoring touched at exactly the two reported points and nowhere else: the A5 bootstrap term (untouched) and the calibration-window note below. Zero golden/simulation baseline edits; no asserted number changed.

Scoring-inertness proof

• PIPELINE_VERSION untouched at 21.
• A5 numbers above; interactions rows kept, so effective_feedback_count is bit-identical.
• The deleted ACEContext.anti_topics / topic_affinities fields had NO reader in the scoring pipeline (v19 removed them; the quarantine loader returned an empty map by construction) — their only remaining readers were two LLM-prompt injections: analysis_rerank.rs (block never fired — live table 0 rows) and the digest prompt, which is now unconditionally the same "None specified" string it always produced, keeping the prompt byte-identical.
• Full suite green with zero baseline adjustments: the scoring simulation + validation suites (inside cargo test --lib) all pass unmodified.
• Three simulation tests were REMOVED, not adjusted — they asserted the ACE anti-topic scoring mechanism deleted in v19 (AD-029) and had passed vacuously ever since: ace_anti_topic_excludes_matching_content (a pure-Rust persona scores python items <0.05 regardless), breakdown_anti_penalty_range (plain range check), penalties_asymmetrically_stronger_than_boosts (its two contexts became literally identical). Removing a vacuous test changes no asserted number.
• Calibration-window drift (forward-looking note, not a change): the autophagy calibration analyzers read interactions over rolling windows. v20b stops NEW implicit rows, so those windows gradually shift toward explicit-only as legacy rows age out. Day-0 output is identical; the drift is organic data-mix change, not a computation change.

Migration Phase 105 (TARGET_VERSION 104 → 105)

DROP TABLE IF EXISTS topic_affinities;
DROP TABLE IF EXISTS anti_topics;
DROP TABLE IF EXISTS activity_patterns;
DROP TABLE IF EXISTS source_preferences;
DROP TABLE IF EXISTS persona_posterior;
DROP TABLE IF EXISTS posterior_snapshots;   -- orphaned by #474
DELETE FROM digested_intelligence WHERE digest_type = 'decision_outcome';

NO DELETE/UPDATE on interactions (A5). Historical Phase 89/90 keep byte-identical behavior via a frozen inline copy of the recompute SQL (the live const was deleted with tracking.rs's learning layer). New migration test test_phase_105_drops_implicit_capture_tables follows the existing wind-back idiom (seeds legacy tables + a decision_outcome digest + an implicit interaction row at v104, re-migrates, asserts dropped/purged/KEPT).

ACTIVATION REQUIREMENT: rebuild BOTH fourda and fourda-engine before anything runs after merge. An old binary opening the schema-105 database is the exact incident class that quarantined the live corpus on 2026-08-16 (old-binary/newer-db → corrupt-db fallback). Migrate and rebuild together.

Zero-reader proof class (separate justification)

autophagy/decision_outcomes.rs was deleted not as implicit capture but as a zero-reader: it wrote decision_outcome digests that no code ever read back (its module doc claimed "results feed back into scoring" — false). Unwound through AutophagyCycleResult (ts-rs binding regenerated by the suite, not hand-edited), digest.rs, autophagy_commands.rs.

Re-sourced consumers (kept features made honest)

• standing_queries_suggestions.rs: suggestions now come from 3+ explicit positive-engagement interactions (the fix(scoring): close the two behavioral inputs that survived the v19 quarantine — PIPELINE_VERSION 20 #476 six-type predicate over interactions, mirroring detect_skill_gaps); 4 tests re-fixtured, incl. an explicit guard that legacy scroll rows never count.
• developer_dna.rs::get_top_engaged_topics: same predicate; return shape preserved; interactions is now an honest count (previously affinity_score as u32, i.e. almost always 0).

Inventory deviations found and handled

• commands.rs (not ace/behavior/commands.rs) held the decay call — as inventoried by content, path corrected.
• personas.rs persona anti_topics feed the KEPT exclusions scoring input — left fully untouched (deleting them would have changed validation numbers).
• The i18n key existed in 13 locale files total (en + 12), not en + 13.
• differential.rs / ace_validation.rs contained three anti-topic fixture sites the inventory's "~8 simulation files" bucket did not enumerate — handled as above.
• source_items.view_count is now write-only (its only reader was the deleted return-visit strength boost). The write was KEPT (not in the inventory's removal scope); flagged as a v20c candidate.
• Stacks source_preferences name-trap honored: zero changes under stacks/.

Verification (all required gates)

Check	Result
cargo test --lib (full, incl. scoring simulation + validation + migration tests)	4380 passed / 0 failed / 8 ignored
cargo clippy --lib -- -D warnings	clean
cargo fmt (targeted, all touched files)	applied
New #[allow(dead_code)]	none
Golden / simulation baselines	zero edits
pnpm run test (FE)	1242/1242 passed (111 files)
pnpm run typecheck	clean
pnpm run i18n:types	regenerated (diff committed)
npx eslint (touched TS)	0 errors (3 pre-existing warnings in untouched DataExportPanel lines)
check-file-sizes.cjs	exit 0, no new errors
check-retired-claims.cjs	OK (after the two disclosed history markers)
ghost-commands.cjs	0 new / 0 stale (both cleared backlog entries removed with their commands)
validate-commands.cjs	consistent at 381 (was 383)
Sacred tests	all four skill-gap tests, anti_topics_no_longer_feed_the_negative_stack, detected_tech_decay_runs_and_does_not_compound, test_engagement_summary_shape — all pass

Known-external on this base (per orchestrator): the NOTICE attribution delta from #450 (fix merging as #486) — not regenerated here to avoid conflicting; resolved by rebase at landing.

DO NOT run the migration or touch data/4da.db until both binaries are rebuilt.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LBY5q2MskiKMkTm6NmB49L