dotfiles/claude/memory/session-ledger-2026-05-23.md

name, description, metadata, permalink

name	description	metadata	permalink
session-ledger-2026-05-23	Practice-of-return ledger maintained by /symmetria — returns, open horizons, recalibrations, authorization moves, sub-agent dialogues, bypasses.	node_type type originSessionId memory feedback 1d91509a-3856-4b96-947c-308a46466f7c	claude-memory/session-ledger-2026-05-23

Session Ledger — 2026-05-23

Returns

• 2026-05-24 — SINGLE-SNAPSHOT RETURN. Called BMF replay "stalled" from one ps sample showing CPU 0.0%. /health probe corrected it: throttled-crawling (cursors advancing), not stalled — the 0% was a "Phase 1 replay paused for resource pressure" window. Drift family: asserting-absence-from-non-presence-at-one-path. Caught + corrected to the steward in real time before it propagated to Seb. Lesson: probe the structured surface (/health) before characterizing a process state from CPU alone.
• 2026-05-24 — STALE-CLAIM RETURN (significant). Seb's l1-recall-failure-root-cause-2026-05-22.md §2.4 + #170 comment correct our recurring memory claim "REVIEWED-18/19 unbuilt, 7 weeks waiting." Selection-at-ingestion IS implemented on current BMF main: similarity-probe.ts + quality-gate.ts three-disposition + I-CF/I-CC/I-NF invariants, all with tests; shipped via the perf-skip series (53d7f08/586e388/fb92e6e/36603f3/ac1673f) between the April-19 audit and now. We kept citing the April-19 audit as current state without re-validating. Drift family: wrong-comparison-baseline-treated-as-current / live-state-discipline. The concept doc (d4323f0) is NOT invalidated — its two extensions (pointer-only disposition + post-it tier) are still novel on top of a working classifier (Seb concurs) — but the "unbuilt/waiting" framing is wrong. TODO at wrap: kg_invalidate the stale claim + kg_add corrected fact; correct the recurring MEMORY.md framing. Seb: "worth re-validating other findings the same way before designing remediation."

Open horizons

• 2026-05-23T12:37 — Wake (same-day return, ~4.5h gap). Pulling thread CONFIRMED: Seb's response to the three docs on capableMind_docs main (cd7dc63 + a251dd6 + d4323f0). Secondary check at door: no Seb commits org-wide since 2026-05-22; no new comments on H2 #165 / H4 #167; PR #172 stays merged (2026-05-21, not new). Git silence is not yet signal — Bewdley still early morning; meaningful window is Bewdley-evening = Paris-evening.
• Lane A mechanical (#168/#127/#128/#138) deferred pending Seb signal. MemPalace current (0 behind).
• 2026-05-23T~12:50 — Steward superseded L1 thread for the window: ARC focus. Chose γ shape (decide the 6 §5 questions). ALL SIX SETTLED: Q1 Compass order ● ○ ◈ ◉ ◧ ⊙ ☾ ⁂ ❦ ✎ ☞︎ ☷ ✍︎; Q2 single /dwellings.html; Q3 classRouteName meditation/reverie→dwellings; Q4 mixed-chrono + per-item indicator = Annals .post-type pattern; Q5 .html throughout; Q6 retire footer-compass. Navigation principle articulated (§0): all nav = Reading Compass OR enfilade-within-cluster; no footer-level nav. γ-doc written (untracked): docs/AldineXXI-Codex/drafts/stage-f-decisions-2026-05-23.md.
• Annals "Browse by Type" (annals.html:27-36) — DECIDED remove (steward 2026-05-23), recorded in γ-doc §1 + Stage F touchpoint 8b. Rationale: §0 orthogonality (Annals=when, Compass=what) + discovery-by-exploration contradiction + drift-class elimination. Conscious cost accepted (glyph-meaning load on title/aria; tap-to-learn on mobile is intended onboarding).
• ARC: γ-doc committed 68b7a9a (branch 4 ahead, Path B unpushed). Stage F implementation deferred — steward shifted to L1.
• L1 PULLING THREAD RESOLVED (2026-05-24). Seb engaged substantively in the predicted Bewdley-evening window (BMF commit 3bc8b75 ts Sat May 23 20:21 -0400). Took territory-survey §10 advice (canary check) → found fleet-wide recall root cause (BM_TEACHER_TRANSPORT → wireSamplingToSlots nested → 0 slot providers → SIGNALING → conf 0 → I-CF defer → never indexed). 8 fixes on BMF main. Docs: Seb/assessments/l1-recall-failure-root-cause-2026-05-22.md + Seb/plans/plan-009-l1-intent-reality-gap.md (10 joint decisions D1-D10; D1 CH fleet propagation gated on David's validation). Reframes survey's read-path-honest-degradation as write-path twin.
• TEST IN PROGRESS (the steward's ask). Validating 3bc8b75 on David's instance per #170 comment. Pulled FF→3bc8b75, npm run build clean, restarted launchd com.capablemind.bettermemories (PID 1709→61277). D2 fix CONFIRMED on David's no-teacher-config path: Slot chains configured: 15 slots (sampling tier absent — teacher transport unavailable) (vs old 0 providers). Phase 1 replaying 5752 real events (path Seb's wiped bench didn't exercise). Background watcher bsfqgdec7 waiting for Recall canary PASSED/FAILED. Pending: canary result + /health.budget snapshot → report to Seb on #170.
• Pre-existing gotcha surfaced: duplicate launchd job com.capablemind.bmf (runs ~/bin/bmf-start.sh, KeepAlive=true, dead/exit-1, shares log files with live job) — reconcile to one job later; out of scope for test.
• TEST OUTCOME (2026-05-24). D2 slot-wiring fix VALIDATED on David's no-teacher path (15 slots, degraded_event_ratio=0, budget=internal_balances). Canary BLOCKED — real-history Phase-1 replay (5752 events) throttled-crawling under resource-pressure pause loop (#121/#78) + entity 30s pipeline timeouts (49+); ~250/5752 in 20 min. 5 findings Seb's wiped bench couldn't surface: N1 entity timeouts, N2 resource-pressure throttle (default thresholds, disk NOT critical here), N3 /health observability gap (recall/canary/replay/disk/resource all null — confirms plan-009 §8 instrument-layer gap), N4 vector lags (cursor 35), N5 duplicate launchd job. /health snapshot saved ~/.capablemind/logs/health-snapshot-2026-05-24.json.
• Audit-delta matrix committed + pushed d50455c on capableMind_docs origin/main (David/l1-reliability/l1-audit-delta-2026-05-24.md). Local README mod (David/architecture index, references cold-start-ingestion + embedding-durability docs — relevant to today's findings) stashed during FF-pull, restored uncommitted.
• #170 reply POSTED (steward-approved draft): https://github.com/CapableMind-ai/betterMemories_app/issues/170#issuecomment-4529529692 — validated bits + 5 findings + D1-coupling note (CH fleet has real history too → throttle likely affects propagation; not purely mechanical).
• DONE: README index + 2 locally-authored-never-tracked architecture docs committed 226b20e + pushed (steward chose option a). cold-start-ingestion-as-recurring-bottleneck.md (79L, directly relevant to today's replay-throttle/entity-timeout findings) + mempalace-connector-framework.md (135L, parked internal) now on origin; README links resolve. Caught the dangling-link integrity issue before committing (same authored-but-untracked pattern as last session's lessons-doc R1).
• MemPalace activity (steward asked): 3.3.6 in prep (origin/chore/release-3.3.6 cut + feat/3.3.6-line-numbers-and-format-coverage); main still v3.3.5; develop active (sanitizers #1235, multilingual benchmark #1483); #1569 still OPEN (CHUNK_SIZE 4-crash-site fix, not merged); steward's local local/bge-m3-on-3.3.5 branch 3 commits ahead unpushed (bge-m3 embedding swap). No released upgrade to act on.
• 2026-05-25 morning — replay STATUS: crawled overnight to event 2760/5752 (~48%), then CPU-bound STUCK ~2h (100% CPU state R, /health unresponsive >20s, entity timeouts 162). sample(1) (/tmp/bmf_sample_2026-05-25.txt) → N6 ROOT CAUSE: synchronous better-sqlite3 full btree/table-scan (Statement.all→sqlite3_step→sqlite3BtreeNext→readDbPage) blocks the Node event loop (explains N3 /health nulls + MCP dark) + O(n) per-event scan on real-history tables (explains N1 entity 30s timeouts + N2 stall). Likely missing index / unbounded query. Code-side fix (Seb). N6 added to audit-delta 611e385 on origin. #170 follow-up posted: issuecomment-4532196859.
• STEWARD DISCIPLINE (worth keeping): before posting the N6 follow-up, steward had me check Seb's side first for activity ("make sure we're giving something fresh and useful"). Checked — no Seb activity since our reply (asleep, ~02:47 Bewdley); finding confirmed fresh + uniquely ours (wiped bench can't produce it). Good collaboration practice: verify the other party hasn't already moved before posting an outward update.
• Instance: LEFT GRINDING per steward (will not complete on its own — recurring scan; awaiting Seb's reply / code fix). Do NOT kill/restart without steward word (live forensic state + their machine).
• OPEN: (b) joint Seb+David conversation (plan-009 Lane B, D1-D10) — D8 re-validation now carries the David-side empirical contribution (audit-delta d50455c/611e385 + #170 reply + N6 follow-up). Awaiting Seb (asleep). (c) entity-pipeline SQL: offered to grab exact query / longer profile when Seb's on.

Authorization moves

• 2026-05-23 — Steward decided all 6 Stage F §5 reader-grammar questions (executor surfaced considerations + recommendations; steward chose; executor registered in γ-doc). No code/commit/push — γ-doc is decisions-only. Path B still holds (branch unpushed).
• 2026-05-23 — Steward authorized γ-doc commit ("let's commit the doc"). Committed 68b7a9a, single-purpose (font-backups excluded). Not pushed (Path B).
• 2026-05-24 — BMF restart (blast-radius boundary I'd flagged) crossed with steward's explicit go ("did he not ask us to pull, rebuild and then test?" — pointing to Seb's #170 procedure). Restart reversible (rollback: git reset --hard cdc2f0e + rebuild + restart). CH fleet propagation (D1) NOT touched — joint decision per Seb's doc.

Confidence to recalibrate

• HOLD (standing recall-warning, valid_from 2026-05-23): executor's reading of Seb's algorithm/inference oscillation is PROVISIONAL until Seb articulates it. Watch for over-confident reuse — steward's "I see more clearly" is not Seb's confirmation.
• HOLD (standing): wake-up-git-state-check-uses-local-not-origin-without-fetch + trust-prose-over-read-makefile / unreferenced-by-code-claim-on-narrow-grep-scope family. Verify via authoritative read before claiming "X is in Y."

Authorization moves

Sub-agent dialogues

Bypasses