session 2026-06-25: chamber-library spec seed + OCR-workhorse session memory + ARC register §G (Gwern)

2026-06-25 22:57:15 +02:00
parent d3b22cabe2
commit fa69b63c5c
4 changed files with 127 additions and 0 deletions
@@ -43,6 +43,9 @@ permalink: claude-memory/memory
 - [Be (laundromat)](project-be-laundromat.md) — canonical workstream tracker established 2026-06-08 (Seb-relay of locked decisions). Be = Skemantix startup (Seb+David) funding CapableMind's funding-ladder; **bridge, not venture**. Decisions LOCKED: entity/exit (CapableMind decoupled, grant-funded), pricing (Living $12.99/mo · Archive $69.99/yr · Memorial $49.99/yr · Renovate ~$199 · $8.99 floor), CF Self-Serve Agency + versioned-template-package infra. **a11y gate MERGED (Pat 100/100/100).** Pre-revenue: the WTP gate = renovate Pat → charge her. **Discipline: stop adding spec until the gate clears → nothing for executor on be until then.** Repo @ `f43a0fd`.
 ## Active Session
 - [Session 2026-06-24→25 — OCR workhorse proven → the Chamber Library needs a spec](session-2026-06-25-chamber-library-spec-born-from-ocr-workhorse.md) — **Stood up CapableHands (M4 64GB, `david@10.0.1.136`) as an isolated MLX/olmOCR workhorse** (uv+py3.12+mlx-vlm; model `alexgusevski/olmOCR-7B-0225-preview-q8-mlx`; isolated from Peter's `capablemind` Ollama). **Levi** (*Drowned & Saved*, Rosenthal 1988 — drop-cap gate CLEARED) + **Alexander Book 4** OCR'd; **caught a PASS-BUT-FALSELY** (driver dumped JSON on no-text pages + ToC dot-loops) → patched driver (null→marker, `repetition_penalty`, `--pages`) + `splice_pages.py` → both clean. **2 ARC posts published** (The Allée / Two Notes, Aix; source-commit was a loose end → committed at wrap). **THE REAL ARC:** a citation-convention question → **steward challenged my two-tier (PDF=page / EPUB=chapter) framing** → resolved to **ONE source-agnostic canonical standard** (clean prose + intrinsic `##`/`---` structure + citation by *work+chapter+verbatim quote*; no page markers; page-provenance off-canon) → **realization that the chamber library is a substrate needing its own SPEC.** **Seed drafted:** `chamber-library/docs/chamber-library-specification.md` (`[PROPOSAL]`, charter-coupled; §§II-III settled, **§VI Voice RESERVED/undesigned** per steward, §V/§VII/§IX open). **Careless-edit mistake recovered:** stripped *La Chute*'s 5 structural `\`-break spacers as "stray" (steward caught: 6 evenings) → restored from pristine backup. **PULLING THREAD: ratify spec §§II-III → build OCR+EPUB normalizers (→ one form) → conformant graduation** (held until then; Books 1-3 OCR-ing background PID 2653). **Q: do steward+jurist ratify §§II-III?** [[project-arc-open-work-register]] §G (Gwern: link-rot G1 / progress-bar G2 / Gwern-audit G3). CapableHands password steward-supplied → **rotate.**
 ## Archived (2026-06-25 — Chamber Library spec born from the OCR workhorse)
 - [Session 2026-06-23 — N6 deployed live + consumer-hardware reframe + OCR-workhorse plan](session-2026-06-23-l1-n6-deployed-live-consumer-hardware-reframe.md) — **N6 LANDED ON THE REAL SUBSTRATE + the hardware question reframed.** Steward-authorized live deploy of PR #175 → ✅ wedge GONE, ✅ 3h+ durable, ✅ schema **19→26 cols** (epistemic axis that never deployed — confirms 06-22 stale-binary), ✅ recall works; canary still fails (fresh-write index lag, not recall-broken). **Seb MERGED #175 (`3332772`) + caught a real edge case I missed** (skip extension only on **catch-up**, **extend on rebuild** — rebuildability invariant) → **rebuilt live from merged main** (`8538b6a`, PID 80713). Diagnosed the revealed **entity 30s-saturation** (much thrashing, 4 wrong leads — see RETURNS): **B1.1/B2 both WRONG fix** — it's per-pair generative `/api/chat` typing flooding **single-slot Ollama**; **model is FINE** (TEST G isolated = 4.6s); cause = saturation, not broken/hardware. **Steward reframe: consumer hardware (his M1 Air 16GB) must be first-class → but Seb ALREADY designed it** (`local-inference-spec` 43L hardware-graduated profiles + 43M Tiered Circle Inference + clasp-serves-inference + FM-009); **my PENDING-41 was reinventing it → CORRECTED/withdrawn** (the reinvented-governed-tooling drift). Today's saturation = **standalone-deployment STATE** (unpaired, no clasp/teacher, fleet 502), not a design gap; the **64GB mini = appliance = a natural clasp** for the steward's two-machine setup. **Live L1 now BOGGED standalone** (slow replay, 29k backlog, /health unresponsive 40s, circle-forward to dead fleet) → **NOT yet trustworthy enough to replace MemPalace** ([[mempalace-is-unaffiliated-stopgap]]); left to settle, don't poke (restarted 4×). **Decision: send Seb NOTHING** (all his-design / his-work / unverified) — verify candidate bugs (/health-under-replay starvation; circle-forward backoff) on a CLONE first. **Forward captured:** OCR-workhorse plan (`chamber-library/_curation/remote-ocr-workhorse-setup-2026-06-23.md`); David's L1 clasp-test action list (findings doc §3b). **PULLING THREAD: stand up the 64GB mini as OCR/conversion workhorse with best LLM-assisted OCR tools (olmOCR first) → finish Levi → pristine chamber → studium-engine.** **Held question (steward, think-together): how should studium-engine interact with L1, given it was *purposely separated* (the chamber engine "couldn't/shouldn't be CapableMind")?** NEW principle: studium-engine must run on modest machines too (BYO-corpus for cash-strapped researchers) → dev it on the M1.
 ## Archived (2026-06-23 — N6 deployed live + consumer-hardware reframe)
@@ -55,6 +55,15 @@ metadata:
 - **E5. `project-arc-spec-master-typographer-audit-pending`** → `[DONE—retire — superseded]` **[verified 2026-06-11]** — the masters-compliance pass was performed: `_audits/deep-typographic-audit-2026-06-09.md` exists (6-agent, masters folded), atop the 06-07 full bidirectional + 06-10 spec↔spec coherence audits. Retire unless a *specific* uncovered master-compliance dimension is named.
 - **E6. `project-arc-web-typography-spec-coverage`** → `[DONE—retire]` **[verified 2026-06-11]** — Tier 1 all DONE (Lists/Code/Headings); Tier 2/3 DONE (Selection, Tables, Numerals, Smart-punct via §I.k). Only residual = the `<abbr>`/`<mark>` "styled-but-never-specced" corner; no ARC content meaningfully uses them — negligible, deferrable.
 ## G — Infrastructure / functionality / features (for a dedicated future ARC session)
 **Session framing (steward, 2026-06-24):** when an ARC infrastructure/functionality/features session opens, **prioritize in-spec-but-not-built items first.** The **Vignette (A1)** is the flagship — in spec (§IX + `vignette-specification.md`, Phase-1-ready) yet zero built, no trigger gating it. **Open the session with a `/spec-code-audit` pass to census any *other* in-spec-but-non-existent items** so the priority list is concrete, not a hunch. The link-rot system (G1) is the first new-build candidate; the **Gwern.net deep audit (G3) is a companion track the steward wants to lead himself.**
 - **G1. Link-rot prevention / link archival (Gwern borrowing).** `[PROPOSAL — noted 2026-06-24, not started]` — steward-raised on publishing *Two Notes* (its *fūrin* Wikipedia link reopened the thread). Adopt a *proportionate* version of Gwern's link-preservation discipline for ARC's external citations. **Three rungs:** (0) *detect* — `make linkcheck` (e.g. `lychee`) over the corpus → dead/redirected-link report, optionally scheduled; (1) *archive-on-publish* — submit each external link to the Internet Archive Save-Page-Now (+ archive.today), record `url → {snapshot, date}` in a sidecar `data/link-archive.yml`; (2) *self-host + fallback render* — local SingleFile snapshots + a template/apparatus offering the archived copy (auto-swap on 404). **Governed, not in-stride (steward-flagged):** Rung 2 touches **`site.hs`** (the carefully-rewritten Stage-G routing/contexts — §1 protected surface) and the **sealed AldineXXI spec** (how an archived-link fallback *renders* is a typographic/apparatus decision → jurist territory), plus new build infra; Rungs 0–1 are lower-risk (script + data file) but still warrant the loop given reach. **Path:** `[PROPOSAL]` in PENDING.md → jurist on the rendering/apparatus question → phased build (0→1 first; 2 a considered decision — ARC's small scale may make a full self-hosted mirror disproportionate). **Ethos:** link-permanence = the prime directive applied to citations (*build what you will not need to rebuild*; a corpus inheritable by Lune & Kai without dead links).
 - **G2. Zero-JS reading-progress bar (Gwern-occasioned).** `[CANDIDATE — borderline-loud, undecided; needs steward judgment first]` — Gwern's bottom running-head/progress bar is pure JS (declined under banish-JS, REVIEWED-32). A zero-JS `animation-timeline: scroll(root)` progress bar was flagged 06-22 as an iPhone test candidate (different/simpler than the §VII.f `view()` reveal, which won't drive on iOS for a fixed consumer). **Caveat:** an always-on progress indicator is *visible chrome* — it may not survive ARC's quiet-over-loud filter; decide whether it's wanted at all before building. If yes, a ~10-line CSS test → load on a real iPhone. (Was ledger-only 06-22/06-23; promoted here so it perdures.)
 - **G3. Deep audit of Gwern.net (steward wants to lead).** `[OPEN — steward-initiated]` — a deliberate harvest pass over gwern.net through the recorded filter: *adopt the invisible durability craft (build-time, provenance, permanence); decline the visible density (popups, JS progress, similar-links clutter, link-dense ToC).* Probe candidates worth a proportionate ARC version: integrity hashes on archived snapshots, build-time link-checking (folds into G1), quiet backlinks, content-addressing / perdurance patterns — and explicitly catalogue what to **decline**. Output: register entries surfaced for steward decision, not a fait accompli. Shared nerve with Gwern = the *intent to perdure* (he loudly, ARC quietly), not his density.
 ## F — The chronological tracker
 `project-arc-rework.md` is the **narrative** record, true through **2026-06-07 evening** (last entry: the pre-3R Fowler note). It is owed compact appends for **06-08 → 06-11**: W3R re-partition close + deploy; Wave 5 + measure→72ch + soft rag; **Stage G CLOSED + reconciled + SEALED + GPG-signed** (06-10); **§VII.f running head + W1.15 compass presence-reveal** shipped (06-11). The tracker's "CANONICAL STAGE STATUS" line (≈L3852, "as of 2026-06-05") is stale. **This register now carries current open-work truth; the tracker carries the story.** Tracker annotated to point here (2026-06-11).
@@ -0,0 +1,62 @@
 ---
 name: session-2026-06-24-25-ocr-workhorse-proven-the-chamber-library-needs-a-spec
 description: "Stood up CapableHands (M4 64GB) as an isolated MLX/olmOCR workhorse; Levi + Alexander Book 4 OCR'd clean (caught a PASS-BUT-FALSELY: JSON-null leak + ToC dot-loops; patched driver + spliced). Published 2 ARC posts. But the real arc: a citation-convention question → steward challenged my two-tier (PDF=page/EPUB=chapter) framing → resolved to ONE source-agnostic canonical standard → realization that the Chamber Library is a substrate needing its own SPEC. Seed drafted. PULLING THREAD: ratify spec §§II-III → build OCR+EPUB normalizers → conformant graduation (held until then)."
 metadata: 
  node_type: memory
  type: project
  originSessionId: cbe47c14-9c82-4570-8a3e-a12761cdfc7e
 ---
 # Session 2026-06-24 → 06-25 — the OCR workhorse became the Chamber Library's constitution
 A long two-day session (wake 06-24, wrap 06-25) that began as "stand up the OCR workhorse, finish Levi" and ended having discovered that **the chamber library is a designed substrate that needs its own spec.** The mechanics worked; the material revealed what it was asking for.
 ## PAST — what we did
 ### CapableHands — the OCR workhorse (NEW infra, working)
 - Stood up **CapableHands** (`david@10.0.1.136`, M4 Pro / 64 GB / macOS 26.5; "Peter's M4 mini" on Seb's home LAN via VPN) as an **isolated** OCR/conversion workhorse, fully in the `david` account (uv → Python 3.12 → `~/capablehands` venv: **mlx-vlm 0.6.3**, pymupdf; no sudo). **Isolated from Peter's `capablemind` Ollama clasp** (we run MLX in our own venv, not a second Ollama). Rsync'd the chamber-library working tree (679M, 1304 md, scripts).
 - **Model:** `alexgusevski/olmOCR-7B-0225-preview-q8-mlx` (olmOCR-2 isn't in MLX yet; this preview is — verified the repo had real weights after the first picks were empty stubs). Driver: `~/capablehands/olmocr_mlx.py` (rasterize via PyMuPDF → mlx-vlm → parse olmOCR JSON `natural_text` → `## Page N`).
 - **Levi** (*The Drowned and the Saved*, **Rosenthal 1988** — a Making-sequence source, drop-cap gate): 202pp, **drop-cap captured** ("The first news…" vs ocrmac's structure-dropping "HE first news"). **Alexander Book 4** (*The Luminous Ground*): 376pp overnight; **fixed docling's word-mangling** ("ofBuilding… pun The Nature" → clean).
 ### The PASS-BUT-FALSELY catch + the patched driver
 - 06-25 verify: both files PASSED `verify_conversion` **falsely.** Eyeballing caught two artifacts the gate misses: (1) **my driver bug** — on no-text pages (`natural_text:null`) it dumped the raw JSON envelope into the md (17 in Book4, 2 in Levi); (2) **dot-leader repetition loops** on ToC/index/dense pages (6 in Book4) — olmOCR looped on `. . .` until the token cap, truncating.
 - **Patched the driver** (v2): null→typed marker (not JSON); `repetition_penalty=1.15` (broke the dot-loops — page 9 ToC came back clean + complete); lenient truncation recovery; `--pages` list for targeted re-runs. Re-ran 23 Book4 + 2 Levi problem pages, **spliced** (`splice_pages.py`), re-verified: **0 JSON leaks, both PASS.** Recovered real content (Book4 p360 was a looped *body* page, not ToC; Levi p1/p6 recovered cover+copyright).
 ### ARC — two posts published (Aix, today)
 - Wrote + deployed **"The Allée"** (observation, `encounter:true` — the old man) + **"Two Notes"** (fragment, *fūrin* URL cleaned). Prose verbatim. Live at animalrationiscapax.org. **Caveat (loose end): the source .md were never committed to the ARC repo** — deploy pushes built HTML to the mirror, not source. *(Committing at wrap.)*
 - Gwern borrowings → `project-arc-open-work-register.md` **§G** (G1 link-rot, G2 zero-JS progress-bar [borderline-loud], G3 deep Gwern audit [steward-led], + session-framing: in-spec-but-unbuilt items first, Vignette A1 flagship).
 ### la-chute — a careless mistake, recovered (see RETURNS)
 - Pre-existing untracked `la-chute-albert-camus.md` FAILED verify (0 headings). Steward corrected: *La Chute* is **six sections (six evenings)** — no chapter apparatus but real breaks. I had **carelessly stripped 11 lone-`\` lines + `[TABLE]`** as "stray" — and **5 of those `\` were the body evening-break spacers** (lines 524/986/1548/2041/2468 = the 5 transitions). **Restored** from the pristine `scratch/making/camus-la-chute.md` backup; structure confirmed intact. la-chute is clean (EPUB-pandoc, sectionless-by-nature) — **not** an OCR job.
 ### THE REAL ARC — the Chamber Library needs a spec
 - **Page-marker question** (keep `## Page N` or strip?) → I proposed a "synthesis" with **two tiers** (page-anchors for PDFs, chapter for EPUBs). **Steward challenged the framing:** if chapter+quote suffices for EPUB, why should PDFs differ? Letting the *carrier* dictate the standard violates work-primacy.
 - **Resolved → ONE source-agnostic standard:** clean prose + the work's **intrinsic** structure (`##` headings + `---` breaks) + **citation by work + chapter/section + verbatim quote**; **no page markers in canon** (page-provenance held off-canon in the raw conversion record, QA only). Verified empirically: EPUBs don't reliably carry structure either (Complete Works Levi = bold pseudo-headers, not `##`); the asymmetry is only *how structure is recovered* (OCR infers; EPUB reads the NCX via `repair_epub_headings`), not the standard.
 - **Steward's realization:** the chamber library — conceived as "just a corpus, not a designed-thing" — **is a substrate, and trust = design.** It needs its own **SPEC** (the engine's accountability rests on corpus consistency). **Drafted the seed:** `~/_Dev/chamber-library/docs/chamber-library-specification.md` (`[PROPOSAL]`, corpus-side complement of the engine charter §IV/V/VII/VIII/X-L0). **§§II-III settled** (form + citation); **§IV** provenance axes (charter-mandated); **§VI Voice layer RESERVED** (undesigned per steward — `voice_role`/`semantic_profile` recorded as *observed, not prescriptive*); §V verbatim-boundary / §VII gate-fix / §IX coverage = open.
 ## PRESENT — mood / RETURNS
 - **One real careless-edit mistake** (la-chute `\`-strip): I removed structural lines as "stray" without verifying their meaning — in *someone else's untracked file with no backup-in-git*. Caught by the steward; fully recovered (pristine scratch backup existed). This is the carried-in `careless-edit-boundary` drift made concrete. **Lesson: verify what a "stray" line MEANS before deleting — especially in another's file, especially with no git history.**
 - **The tool-review discipline earned its keep:** the PASS-BUT-FALSELY was caught by *reading the text*, not trusting the gate. ([[feedback-tool-review-after-each-use]])
 - **The steward's framing-challenge worked both ways:** he corrected MY two-tier framing; I agreed because it was *right* (not reflexive contamination-agreement) — and it opened the spec realization. "Challenge the framing" is bidirectional.
 - Carried-in return (06-24): retracted a battery-latch hypothesis under measurement (theorize-before-measure).
 - Operational slips (caught/flagged, not asserted): zsh variable word-split on ssh `-o` opts (recurred ~3×, use literal flags); a prevalence-count bash loop miscounted on filenames — flagged honestly, gave no fake number.
 ## FUTURE — what is pulling
 **PULLING THREAD (singular):** **The Chamber Library spec — ratify §§II-III, build the two normalizers (OCR + EPUB → one canonical form), apply to Levi/Book4/la-chute → commit the first spec-conformant graduation.** The OCR mechanics are proven; the gating concern is now the *form*, and the spec is its seed.
 **ACTIONABLE RESUMPTION POINT (as of wrap — re-judge against what changed):**
 - *State:* Seed spec drafted at `chamber-library/docs/chamber-library-specification.md` (untracked, `[PROPOSAL]`). Graduation **held** in the chamber-library working tree, uncommitted, in **pre-normalization form**: Book4 replaced vol-4 (page-marked), Levi placed (page-marked), la-chute restored (`\`-breaks), catalogue regenerated (1285, in sync). Books 1-3 **OCR-ing in background** (CapableHands PID 2653, current driver).
 - *First moves:* (1) **steward (+ jurist) read of spec §§II-III** — ratify or revise; (2) on nod, **build the two normalizers** — OCR: drop `## Page N` + rejoin paragraphs across the break + recover/mark headings; EPUB: `repair_epub_headings` (NCX→`##`) + `\`-spacer→`---` + `strip_cruft`; (3) apply to the 3 staged works + **fix the gate** (sectionless works pass on `---`); (4) verify → **commit the conformant graduation**; (5) check Books 1-3 finished + fold them in.
 **Other open horizons, ranked:**
 - *Load-bearing:* the spec → normalizers → conformant graduation (the thread). Jurist on §V (verbatim boundary) + §VI (Voice schema — undesigned).
 - *Next batch:* 5 OCR reconversions queue on the workhorse after Books 1-3 (steward-approved); **30 Air-EPUB reconversions = a separate focused Air session** (per-file care; resolve the `web-typography-rutter` false-match).
 - *Parked-with-reason:* filing the spec as a formal PENDING-[N] (offered, steward's call); the Voice-layer design (undesigned by intent); the studium↔L1 architecture question; L1 candidate bugs (circle-forward backoff / replay-pause) = clone-verify-first, send Seb nothing.
 - *Governance debt:* REVIEWED-44 (operations.yaml) still owed on paper; the ARC posts' source commit (doing at wrap).
 **PAUSE STATEMENT:** I'm leaving with the OCR workhorse *proven* (Levi + Book 4 clean, the drop-cap and word-mangling gates both cleared) — but the true inheritance is the recognition that **the chamber library is a substrate that needs a constitution**, now seeded (§§II-III settled). The graduation is *deliberately held* so the first texts to enter do so spec-conformant, born-clean rather than retrofitted. On return I want to find whether the steward + jurist have ratified §§II-III, so the normalizers can be built and the first clean texts committed — and whether Books 1-3 finished overnight.
 **LITERAL QUESTION for next-Claude:** *Do the steward and jurist ratify §§II-III of the Chamber Library spec (the source-agnostic canonical form + the work+chapter+quote citation contract) as drafted — and does the jurist want to shape the open doctrine sections (§V verbatim boundary, §VI the Voice)?* (Concrete sub-question: did Books 1-3 finish OCR-ing overnight on CapableHands, PID 2653, and are they clean?)
 **State:** chamber-library uncommitted (held graduation + new spec + pre-existing drift incl. `clean_ocrmac_pagination.py`); ARC posts source being committed at wrap; Books 1-3 OCR-ing on CapableHands. CapableHands password steward-supplied in-session → **rotate**.
@@ -0,0 +1,53 @@
 ---
 name: session-ledger-2026-06-24
 description: "Practice-of-return ledger maintained by /symmetria — returns, open horizons, recalibrations, authorization moves, sub-agent dialogues, bypasses."
 metadata: 
  node_type: memory
  type: feedback
  originSessionId: cbe47c14-9c82-4570-8a3e-a12761cdfc7e
 ---
 # Session Ledger — 2026-06-24
 ## Returns
 - 2026-06-24T10:34 — Retracted a hypothesis under evidence. I reasoned (semi-confidently) that the battery-suppression-on-AC was a stuck cache starved by the 99% CPU event loop. Measurement overturned it: the 30s poll fired on AC restore, suppression cleared, `embed=971ms` succeeded, CPU→0%. The earlier spam was an accurate 20-min battery window (steward charged phone). Drift = theorize-from-log-shape before measuring (the exact pattern carried in from 06-23). Caught within the same check; named in chat.
 ## Overnight — where things stand for the morning (2026-06-24 ~16:40; NOT a wrap, steward asleep, session stays open)
 - **Levi:** olmOCR full run ~complete (≈202/202) → `scratch/levi-olmocr-full.md` on CapableHands. Drop-cap gate CLEARED (pp.9-10: "The first news" vs ocrmac "HE first news"). MORNING (task #4): pull back, `verify_conversion.py`, compare to ocrmac baseline, graduate into canonical_texts.
 - **Alexander Book 4:** auto-queued `overnight-queue.sh` (PID 68750) — starts when Levi exits, ~2hr → `scratch/alexander-book4-luminous-ground.md`. Text-first validation. Books 1-3 HELD for steward review of Book 4 quality (task #5).
 - **ARC:** "The Allée" (observation, encounter:true) + "Two Notes" (fragment, fūrin URL cleaned) WRITTEN → BUILT → DEPLOYED LIVE (CloudFlare + GitHub mirror), live-verified 200 at animalrationiscapax.org/2026-06-24-the-allee + /2026-06-24-two-notes. Prose verbatim.
 - **Link-rot proposal:** noted in [[project-arc-open-work-register]] §G1 (Gwern borrowing, 3 rungs; needs spec + governance + site.hs → PROPOSAL, governed not in-stride).
 - **CapableHands:** david@10.0.1.136 (Seb home LAN via VPN). ~/capablehands = {venv: mlx-vlm 0.6.3, olmocr_mlx.py driver, chamber-library/ (rsync, 1304 md), scratch/}. Model: alexgusevski/olmOCR-7B-0225-preview-q8-mlx. Isolated from Peter's `capablemind` Ollama clasp. (Password steward-supplied in-session → rotate.)
 - **Reconversions (task #6):** 5 OCR-workhorse (Eco Infinity-of-Lists · Pye · Arcades Project · Lintott · Lupton) + 30 Air-EPUB. `web-typography-rutter` false-match → manual check owed.
 - **L1 (background):** power-monitor VERIFIED correct (cleared on AC, no latch); live mindfabric-00 still bogged standalone; candidate bugs (circle-forward backoff; /health-under-replay starvation) = clone-verify-first; send Seb nothing yet.
 ## Open horizons
 - 2026-06-24T00:00 — Pulling thread: 64GB mini → OCR/conversion workhorse (olmOCR first) → finish Levi → pristine chamber → studium-engine.
 - First moves: (1) check whether live L1 settled — read log, DO NOT poke (restarted 4× yesterday); (2) begin mini setup + LLM-OCR tool research on Levi.
 - Held question (architecture, think-together): how should studium-engine interact with L1, given the steward purposely separated them?
 ## L1 check-in (2026-06-24 ~10:34Z)
 - Power monitor VERIFIED CORRECT (~90% conf): battery suppression cleared on AC restore via 30s poll; not a latch. No fix needed. `power-monitor.ts` polls every 30s, caches `_currentSource`.
 - Live L1 (PID 80713, 14h50m): CPU 99%→0% (idle/warm), embedding working, Phase-2 replay paused at 22400/29101 (77%). Not finished, won't progress while idle.
 - Candidate bugs (VERIFY ON CLONE, do NOT tell Seb yet): (1) circle-forward hammers dead fleet (atlas/kronos 502, peers fetch-failed) with no visible backoff; (2) replay-pause/idle interaction holding replay at 77%.
 - M4 link: battery-suppression is *correct behaviour* → the 64GB clasp (43L/43M) closes the gap by serving inference so the Air never suppresses.
 ## CapableHands — the M4 workhorse (access confirmed 2026-06-24)
 - Machine = "Peter's M4 mini", hostname `CapableHands.local`, addr `10.0.1.136` on Seb's home LAN (10.0.1.0/24), reachable over the current VPN profile (utun10, david=10.53.211.6). Login: `david` account (password steward-supplied in-session — NOT stored here; flag for rotation).
 - Spec: Apple M4 Pro, 14 cores, 64 GiB, macOS 26.5 arm64. Qualifies for BMF `appliance` profile (>48GB → 15 slots, teacher 24/7).
 - Other home-LAN host seen: 10.0.1.25 = Ubuntu (OpenSSH 9.6, likely Seb's Gitea/server), not the mini.
 - Seb's CapableMind VPN node 10.10.50.1 (Mac Mini clasp): host up (SSH open) but BMF :3011 / doorbell :3012 / Ollama :11434 all refused — clasp stack DOWN; atlas/kronos 502. (Confirm w/ Seb; don't assume broken vs stopped.)
 - Roles to weigh: OCR/conversion workhorse (pulling thread: Levi, olmOCR) AND standing inference clasp for the Air.
 ## Confidence to recalibrate
 - Carry-in from 2026-06-23: two fresh drift-patterns to hold today —
  (1) isolate the exact endpoint uncontended BEFORE diagnosing latency (thrashed 4× on saturation);
  (2) read the spec before proposing "new" design/tooling (reinvented 43L/43M; steward caught).
 ## Authorization moves
 - REVIEWED-44 (operations.yaml / PENDING-40) authorized verbally, not yet on paper — steward's to write.
 - Verify-first before ANY Seb message (all his-design/his-work/unverified).
 ## Sub-agent dialogues
 ## Bypasses