dotfiles/claude/memory/session-ledger-2026-04-18.md

---
name: Session Ledger 2026-04-18
description: Practice-of-return ledger maintained by /symmetria — returns, open horizons, recalibrations, authorization moves, sub-agent dialogues, bypasses.
type: feedback
originSessionId: e55939da-2fa7-4e18-a2e5-8f8fd3a8c607
---
# Session Ledger — 2026-04-18

## Returns

- **~10:25** — I named the thread-shift as if the session had re-pointed; steward corrected: the governance-substrate exchange was "just on my mind this morning," a thread caught before it slipped, not a redirection. Original pulling thread (Ethics of the Reply publication — Hakyll vs. typography lever) is restored. Name: *when the steward catches a thread in peripheral time, that's not a re-pointing of the session; the held thread remains the held thread.*

- **~10:45** — Was about to surface a URL-structure judgment call on Hakyll versioning (canonical URLs `/foo.html` vs spec `/essays/foo/`). Before drafting the question, checked yesterday's staging README — line 63 already commits to canonical-unchanged + prior-versions-at-path-segment. *The question was answered yesterday; I almost duplicated the work by re-asking it.* Name: *before surfacing a judgment call, check whether prior work has already resolved it. The staging README is the canonical record of yesterday's decisions.*

- **~15:30** — Ran `make rebuild` to verify chamber voices archival + colophon cleanup, without knowing that `make build` auto-deploys to CloudFlare Pages + GitHub mirror. The archival + cleanup went live BEFORE I committed them locally. The deployed state matches what the steward authorized (no rollback needed, content is correct), but the order was wrong — commits should precede deploy. Name: *before running any `make` target that could touch shared infrastructure, grep the Makefile for what it actually does. `make build` and `make rebuild` both cascade into `deploy-all`. There is no build-without-deploy target. For local-only verification, need to either add a new target or use `make serve` (which builds+serves but may also deploy — unverified).* Downstream impact: Ethics I & II go-live will also auto-deploy on `make build`. Surfaced to steward for explicit re-authorization before proceeding.

- **~evening** — Cited `#10 sequential dispatch` as a live constraint on vault ingestion duration, without checking its status. `gh issue view 10` showed CLOSED 2026-03-24 (~3 weeks ago). Also quoted `eta_minutes: 222` from connector_status as a time estimate, without knowing that ingestion rate is bursty (per steward: *"the ETA is not a good indicator because the rate will change with the burst ingestion pattern that seems to take hold"*). Name: *MEMORY.md captures state at a point in time; issue statuses and system behaviors observed weeks ago may no longer hold. Before citing an issue number as a constraint, `gh issue view`; before citing a BMF-reported metric as a predictor, check whether the observation pattern for that metric is already documented (memory + gh). Search before claiming; search before citing.* Promoted the bursty-ETA observation from incidental session-log note (session-2026-03-21 line 69) to canonical `feedback-bmf-ingestion-rate.md`.

## Open horizons

- **Path to publication — two-lever question** (carried from 2026-04-17). Hakyll infrastructure vs. typography application skill. Steward's call; do not pre-decide.
- **Drift patterns carried from yesterday's KG**:
  - `register-import-without-checking-target-context` — ARC voice ≠ project-notes voice. Active if we touch ARC prose.
  - `fact-fabrication-instead-of-let-me-check` — MemPalace protocol step 3. Active if we touch proper nouns or facts I'm not certain of.
- **Seb has shipped BMF Settings Phase 1 + 2** (plans 006 Phase 1/2, plans 007/008 drafted, pod.io analysis filed) while we were away. Sits beside the thread — may surface questions.

- **cm-world circle join COMPLETE** (evening 2026-04-18). After wrapper edit + kickstart + re-join, full circle participation is working. **Canonical identity: `cm-instance-6a3e981e01613a8c`** (member, tier: default, member_base_url: `http://10.53.211.6:3011`, joined 2026-04-18T16:13:35). Roster visible (5 members including clasp); `/v1/circles/:id/members` on new binary returns data (was 404 on old). Two items to flag Seb next session:
  - **Ghost member entry** — `cm-instance-b7fab457a1d3c9d7` from the first (stale-binary) join is still in the clasp roster. Orphaned: no local BMF owns it, can't `/leave`. Needs server-side cleanup on Seb's end.
  - **Handle routing not wired** — `GET /v1/circles/cm-world/members` returns NOT_FOUND; only UUID `50d99cc2-2860-48e6-b37e-c6c3686106e6` works. Seb's verification command assumed handle works.
- **Wrapper edits shipped** — `~/bin/bmf-start.sh` line 18 token → `bm_key_testkey1234567890` + added `export BM_DATA_DIR="$HOME/_Dev/BetterMemories.io/data"`. Dotfiles copy at `~/dotfiles/bin/bmf-start.sh` got the BM_DATA_DIR line too (token stays `bm_key_REDACTED` — established pattern). Steward to commit dotfiles change at wrap-up.
- **Old data orphaned** — pre-existing logchain at `~/_Dev/BetterMemories.io/data/logchain/` (Mar 6, pre-SQLite migration) is not read by the sqlite-lance binary. New binary created fresh SQLite security records, hence the new instance_id. Old logchain files are dead weight — can be archived/removed at some future cleanup. Not blocking.

- **CORRECTION (later in evening) — the "orphan" framing was wrong; real data lives at canonical path** — after BMF stopped cleanly, inspection revealed dual-write: our `BM_DATA_DIR` pin to `~/_Dev/BetterMemories.io/data/` diverted new-binary writes to the repo path, but the REAL historical data has always lived at `~/.capablemind/data/mindfabric-00/` and the new binary still wrote to it (per-module resolver bypass or ADR-021 fallback touching it at startup). Evidence:
  - REPO `/sqlite/` (our pinned path): entity 233k, temporal 139k, vector 61k — small, today-only, holds the `cm-instance-6a3e981e01613a8c` circle membership
  - CANONICAL `/sqlite/` (real home): entity **2.1 MB**, temporal **5.7 MB**, vector **2.1 MB** — substantive, from prior vault work (Apr 15 recall validation + subsequent incomplete ingestion steward referenced)
- **Lesson**: Seb's ADR-021 fallback behavior + per-module path logic means BM_DATA_DIR is NOT a simple monolithic override. Some modules respect it; some read legacy canonical paths. Don't assume one env var redirects all data. This is worth flagging to Seb as a `[HARDENING]` — "BM_DATA_DIR should be authoritative for all data, no backdoor paths."
- **Decision executed — clean consolidation complete** — archived both: `~/.capablemind/data/mindfabric-00.archive-20260418-incomplete` + `~/_Dev/BetterMemories.io/data.archive-20260418-rogue-pin`. Wrapper (live + dotfiles) now: `BM_DATA_DIR=$HOME/.capablemind/data/mindfabric-00` + `BM_INSTANCE_NAME=mindfabric-00`. BMF bootstrapped on PID 55326, healthy in 1s, fresh 4.1k security.sqlite3 created. **Single source of truth confirmed**: post-bootstrap, `~/_Dev/BetterMemories.io/data/` was NOT re-created by BMF — the combination of explicit `BM_INSTANCE_NAME` + archived legacy fallback targets stopped the per-module path-drift. This is the mitigation for the dual-write bug.
- **Canonical identity now: `cm-instance-7d2e19a417d694c6`** (joined cm-world 2026-04-18T18:17:13 via re-run of join script). **Ghost tally for Seb**: `cm-instance-6a3e981e01613a8c` (now orphaned). `cm-instance-b7fab457a1d3c9d7` already cleaned by Seb earlier.
- **Vault ingestion STARTED** — batch `import_1_mo4nvvyi`, `connector_status obsidian` = `running`, **9,836 total items estimated**. First ingestion on the new stack (SQLite+LanceDB main with post-migration perf fixes). Monitoring at steward's cadence. BMF launchd-managed, survives sleep/restart. Logs at `~/.capablemind/logs/bmf.stdout.log`.

- **Entity-relationships stage bottleneck diagnosed during ingestion** — at 550/9836 processed, rate had dropped 44 → 36 → 13 notes/min (steady downward, not bursty). Module health showed `entity` with 165 errors (every other module: 0-1). Log signatures: `[entity] Pipeline error (deferrable): entity pipeline timeout after 30000ms`, `[entity:pipeline] SLOW stages: resolve=30252ms, relationships=840082ms, persist=15ms (8 entities)` — the **relationships stage took 14 minutes** on one batch, 18 min on another (1080082ms / 9 entities). Entity events deferred on timeout, cursor held, retried — no data loss but significant backpressure. Hypothesis: relationships stage scales with graph density; as ingestion proceeds, batch duration may grow rather than shrink. Worth filing a GH issue if none exists. Vector subsystem is fine (`[vector:hybrid] bm25=10 vec=5 search=342ms`). Distinct from the bursty-ETA pattern in `feedback-bmf-ingestion-rate.md` — that was rate variance; this is persistent downward drift.

- **Circle-forward noise (non-blocking but present during ingestion)** — repeated `[circle-forward] failed to cm-instance-711269556f1e32ca: fetch failed` (Seb's BMF unreachable — may be offline or IP shifted); `POST https://kronos.capablemind.ai/v1/observe returned 504` and same for `atlas.capablemind.ai` (cloud endpoints Gateway Timeout). Local ingestion unblocked — circle-forward is async — but adds log noise and some CPU.

- **Symmetria pulse 2026-04-18 evening before diagnostic GH search** — alignment check after the session's error pattern. Concrete discipline for the entity-relationships diagnostic work: (1) search GH thoroughly before claiming no existing issue; (2) mark "scales with graph density" as hypothesis not claim — symptoms observed, mechanism not verified; (3) include verbatim log signatures so Seb can reproduce; (4) size artifact to one finding, not a treatise. Contamination watch: don't frame hypothesis as claim; treat "NOT failing" as `connector.failed=0` observation, not general health assertion. Alignment: aligned with these adjustments.

- **Entity-relationships diagnostic formalized** — GH search found #65 ("BMF untested for bulk ingestion on single-user laptop — compound failure pattern") as the right umbrella; not a new issue. Posted detailed comment: https://github.com/CapableMind-ai/betterMemories_app/issues/65#issuecomment-4274470826. Companion thinking doc shipped to `docs/thinking/David/l1-reliability/l1-ingestion-diagnostic-2026-04-18.md` as successor to the 2026-03-25 diagnostic, committed and pushed as `d70437c` on CapableMind-AI main. Both reference the 14–18min per-batch relationships stage timings, module health snapshot (entity error_count=165 vs 0–1 elsewhere), and unverified hypotheses. Steward leaving ingestion running overnight — rate stabilization (steady floor) vs continued drop will distinguish density-scaling hypothesis from steady-state-floor hypothesis.

- **Upstream pickup during rebase — Seb shipped BBF (Back Office Framework) work** — during `git pull --rebase` on CapableMind-AI, picked up two Seb commits: `5ea7f5d BBF spec v0.2: peer-infra model + enhancement-stack + Back Office reshape` and `241463a BBF spec reorg: specs/bbf/core + instances`. New body of work visible in `specs/bbf/`. Worth context-grounding for next session; BBF is adjacent to / building on Seb's recent circle-networking + plan-007/008 staff-keys work on betterMemories_app.
- Tunnel `10.53.211.6` is steward's Firewalla-or-equivalent VPN to Seb's LAN (confirmed 2026-04-18 evening). Script file from Seb's Messages: `/Users/davidglidden/Library/Messages/Attachments/77/07/44D97289-D079-43BD-A09C-63FF3635A43D/join-cm-world.sh` (not copied into our repo — it landed on main as `47be814` and is now at `scripts/join-cm-world.sh` after branch switch).
- Second clone at `~/bettermemories/` was created fresh by the script (main @ `decbdee`, built dist/, no BMF started due to port conflict). Can be deleted or kept as isolated build env.

## Confidence to recalibrate

- **Skills-I-co-designed** contamination carried forward.
- **Synthesis before exploration**: the steward's opening was `good morning /wake-up` — no directive yet. Wait for the lever question or a different opening. Do not volunteer implementation.

## Authorization moves

- **Ground-up model horizon planted in MemPalace** — steward authorized the seed-watering practice. Drawer `drawer_capablemind-thinking_seeds_afe475b555e5bbe2` filed with steward's verbatim intention + practice quotes. KG established: entity `ground-up-model-horizon` with 7 facts (status, surfaced_via Greenblatt article, two gaps, connects_to contamination-problem, davidglidden holds-intention, claude-code watches-horizon). Watering convention: new relevant drawer + `<topic> → waters → ground-up-model-horizon` KG fact on surface. Not a task — peripheral vision discipline.
- **First watering filed** — jurist-steward riff on architectural alternatives + Claude Code input (two sharpenings, one addition, one political angle, one calibration). Drawer `drawer_capablemind-thinking_seeds_e441d121ac050cb3`. Six KG facts added (substrate lessons, entry point, interpretability direction, political dimension, tiered-architecture serves horizon).
- **Second watering filed + three L2 directions added to open-tasks** — governance-substrate exchange produced three directions for when L2 resumes. Split authorized by steward: directions 1 (governance audit) + 2 (governance pattern KG) + small CD-03 exercise added to `open-tasks.md` Priority 1 section with pointer to seed; direction 3 (counterfactual instance on open-weight) held as horizon in seed, not promoted to open-tasks (not near-term). Drawer + 9 KG facts including `L2-architecture → mitigates-but-does-not-solve → contamination-problem` and `L2-parked-period → produced → scaffold-for-L2-operability`.

- **Ethics of the Reply I & II v2 go-live complete (3af43ca + deployed)** — first exercise of ADR-005 versioning regime in production. v1 snapshots moved from `_staging/versioning-prep/v1-snapshots/` to `content/posts/` via git rename (preserving history); v2 content from iA Writer translations overwrote canonical filenames. Local build verified before deploy (discipline maintained after earlier auto-deploy surprise). Both v2 canonical URLs render breadcrumb ("Begun Barcelona 2025-06-17 · Revised Barcelona 2026-02-08") and both v1 prior-version URLs render top banner ("You are reading a prior version... · Read current"). Staging README updated to historical note + pointer to remaining III-V + sequence-index work. Yesterday's pulling thread is resolved for I & II; remaining from it are deferred items (vignette, typography skill) and next-publication-session items (III-V + sequence-index). Five commits ahead of origin awaiting push authorization.

- **Dateline dedup fix shipped (e292593 + deployed)** — steward caught visible duplicate on live page: "Barcelona June 2025" appeared twice (inline italic from markdown + mono from template `.dateline` apparatus). Root cause: 2026-04-17 staging decision to include `*Barcelona, [Month Year]*` as inline paragraph (matched PDF) conflicted with ARC's layered apparatus (template renders dateline from frontmatter). Fix: removed inline paragraph from both v2 essay markdowns; staging README updated with explicit lesson for III-V translation ("do not include the inline dateline — template generates it"). Answered steward's serif-vs-mono version-stamp question: **serif is correct by design** — `.dateline` is mono (apparatus register) and `.version-stamp` is serif (editorial/bibliographic register) deliberately, per `_post.scss` and `_utilities.scss` comments. Two different kinds of thing.

- **Typography register — Pattern 3 adopted (2026-04-18)** — after investigation surfaced that the sans register was essentially vestigial (used only in two enfilade navs + an unused heading-base mixin), steward chose Pattern 3 (role-assigned families: serif content / sans wayfinding / mono apparatus) with accessibility as the load-bearing justification — family-change differentiation is more robust under visual impairment than within-family variation, and the steward (myopic, progressives) thinks of readers who need visual support. Decision canonically captured in `project-arc-rework.md` under the style-guide thread: Pattern 3 as the structural system, current font stack (EB Garamond + IBM Plex Sans + IBM Plex Mono) retained pending steward consideration, font-swap candidates surveyed (iA Writer Mono as lowest-regret mono swap; Alegreya Sans / Ysabeau as open-source humanist sans alternatives; Garamond Premier Pro + Cronos Pro as Adobe CC path; Fedra / Scala / Adelle / Brill as paid superfamily options). Counter-arguments held but not overriding (voice tension between Garamond humanism and Plex industrial register; architecture vs. continuous flow; designing-for-self bias). Cleanup implied but not yet executed: delete dead `heading-base` mixin; audit remaining nav surfaces for register consistency. Steward's canonical articulation preserved verbatim: *"serif for the body, sans for nav, and mono for apparatus seems to make things clear."*

- **Listing duplication fixed + nav elements to sans (13757a5 + deployed)** — steward caught that both versions of each essay appeared in `/essays.html` listing (2 essays × 2 versions = 4 entries). Root cause: `loadAll "content/posts/*.md"` matches both current and prior-version files; both have `class: essay`; both appeared. First fix attempt used Hakyll's pattern composition (`pattern .&&. complement priorVersionPattern`) and DIDN'T work — the glob/regex composition in loadAll pipeline behaves differently from expected. Fallback to Haskell item-level filter (`excludePriorVersions` helper using `parseVersionedFilename`) WORKED. Applied to all 7 listing aggregations. **Critical gotcha also caught**: editing `site.hs` alone doesn't trigger a Stack rebuild of the binary — must run `stack build` before `stack exec` to pick up changes. Name: *when site.hs changes don't appear to take effect, check that the binary was rebuilt (stack build) before blaming Hakyll cache.* Concurrently executed steward's nav-to-sans request (Pattern 3 cleanup): `.post-nav`, `.post-header .breadcrumb`, `.tags`/`.tag` all now use `$sans-family`. Reading Compass documented as a **Pattern 3 exception** (iconic Unicode symbols, not lexical nav — family-change doesn't differentiate symbol rendering). Inline comments in `_post.scss` and `_compass.scss` so the decisions read as deliberate, not drift. Pattern 3 is now *operative* in the codebase, not just declared in memory. 7 commits ahead of origin.

- **Breadcrumb made navigable + version-stamp register shifted (7540388 + deployed)** — steward noticed after the listing fix: *"Now I can't access the previous versions... should that be in the begun...revised area?"* My MVP had deferred the Prior Versions link (spec §7 item 7) without realizing the deferral left prior versions undiscoverable. Steward's instinct ran ahead of my spec interpretation: the breadcrumb *itself* is the access surface — each non-current date is a link — cleaner than the spec's separate "Prior versions" text. Implemented: every date in the breadcrumb except the last links to `/<slug>/v/<date>/`; scales to N revisions without code change. Steward's second question in the same exchange: *"this being nav it should be in sans maybe?"* Correct — version-stamp shifted from serif (editorial register) to sans (wayfinding register) per Pattern 3 consistency. Colophon edge case handled: pages outside `content/posts/` detected via `mslug = Nothing` and render plain (no broken links). Symmetria return logged: *MVP boundaries worth re-examining under steward's eye — my "deferred per spec" stance looked principled but was under-thought; the steward's design ran ahead of mine.* 8 commits ahead of origin. 5+-revisions threshold flagged for future switch to per-piece version-index page (spec §7 item 7 properly implemented at that point).

- **Letter-spacing harmonization + Reading Compass note + push (2332db9 + both remotes)** — steward noticed version-stamp looked slightly different-sized than post-header breadcrumb despite both being 0.875rem; diagnosed as residual `letter-spacing: 0.02em` from the old serif editorial register. Dropped at sans; both nav surfaces now harmonized visually. Reading Compass optical rebalancing (steward's observation: *"the reading compass glyphs are meant to be adjusted to each other to be optically equivalent sizes, but I never got that working"*) noted as deferred cosmetic task in `project-arc-rework.md` with three resolution paths (custom SVG / multi-device iteration / accept platform variance). **9 commits pushed to origin (git.skemantix.com) + github-backup mirror** — `42fbddd..2332db9`. Pattern 3 is now fully operative: code + deployed + pushed. All pre-commit checks passed cleanly throughout the session — no `--no-verify` bypasses since the hook fix landed.

- **Hakyll versioning infrastructure shipped (bb9e7eb)** — spec §7 items 1-5 implemented. `site.hs` +166 lines (Revision type, parseVersionedFilename, versionedPostRoute, renderCurrentBreadcrumb, renderPriorBreadcrumb, three new context fields); `templates/post.html` +7 lines (prior banner + breadcrumb after dateline). Uses `fromRegex` not glob — Hakyll glob doesn't support `[0-9]` char classes. Conservative policy adopted (renders only when `revisions:` or `edition_label:` present) to avoid stamping 170+ existing pieces without explicit authorization. First ARC commit without `--no-verify` since the hook fix landed. Verified: colophon renders new breadcrumb correctly; Ethics v1 correctly does not.
- **Style guide scope expanded** — filed as named thread in `project-arc-rework.md` with explicit dependency link to the projected typography application skill (layer 3 expressive decisions need style guide's declared positions as rule set; skill enforces the guide, guide defines what the skill enforces). Practice-first audit recommended over theory-first rewrite.

- **Launchd BMF architecture discovered (evening 2026-04-18)** — `launchctl list` shows `com.capablemind.bmf` (PID 1484) is the ACTIVE one, not `com.capablemind.bettermemories` (exit 1, likely loses port race). Active plist runs wrapper `~/bin/bmf-start.sh` which: sources `~/.capablemind/env` for `ANTHROPIC_API_KEY`; sets `BM_PORT=3011`, `BM_TEACHER_TRANSPORT=api`, `BMF_MCP_TOKEN=bm_key_7feb421ef8fd45e8`, `OLLAMA_HOST=http://127.0.0.1:11434`; `cd ~/_Dev/BetterMemories.io && exec node dist/index.js`. Full env from `ps eww`: also `BM_LOCAL_MODEL=qwen3.5:4b`, `BM_STORAGE_BACKEND=sqlite-lance`, `BM_TEAM_PSK=7727ce…`. Data dir is CWD-relative (logchain at `~/_Dev/BetterMemories.io/data/logchain/`, from Mar 6 — pre-ADR-021 default). Both plist files still exist in `~/Library/LaunchAgents/` — the unused `com.capablemind.bettermemories.plist` is cruft (and duplicates the port-binding intent). Cleanup opportunity: remove the unused plist.

- **BMF branch divergence discovered (evening 2026-04-18)** — `origin/trial/storage-sqlite-lance` (HEAD `3ea031c`) was merged into `origin/main` via `adf4743 Merge trial/storage-sqlite-lance into main`, and main is now 25 commits ahead with the circle-networking features Seb has been building: `24e0186 peer-mesh roster propagation + door pass issuance`, `bb390a2 Keystone forwards observations to circle peers`, `eb4cbd7 auto-establish pairings between circle members on join`, `ce34d7f createCircle accepts clasp_base_url`, `e52c222 roster broadcast authenticates via BMF_MCP_TOKEN`, `ce2ae30 /v1/circles/:id/leave bumps roster + broadcasts on clasp`, `da6bb5f GET /v1/circles/:id/members`, `47be814 one-command onboarding for cm-world circle members` (the script Seb forwarded), `8cf727d bulk-ingest scripts for Atlas docs + Kronos GH issues`, `bed8107 entity pipeline skips slots for cross-circle forwarded events`, `5b14db7 BM_CPU_PAUSE_THRESHOLD env for replay-coordinator`. Plus fixes #124 #133 #134 #135 #136 #149 #160 landed on main. `origin/main` HEAD = `5b14db7`. **Trial branch is effectively stale**; the "latest build" now means main. Our local `~/_Dev/BetterMemories.io` is on trial branch; need to checkout main + pull + rebuild before running join script. Memory-file state at time of discovery still said "branch `trial/storage-sqlite-lance`, 30 commits ahead of main" — MEMORY.md update needed.

## Sub-agent dialogues

## Bypasses