INSIGHTS

Ephemeral capture buffer — running notes, observations, and claims for the user. Append
freely during sessions; this is the fast, messy layer. Settled items get promoted to
evergreen wiki pages at the weekly review (see CLAUDE.md → Promotion).

This repo uses only INSIGHTS.md (everything here is user-facing). The
INSIGHTS/CLAIMS split lives in the code repo, where agent- vs user-facing audiences
differ.

---

Archived: the 21–23 June capture entries were moved to
_done/Jun23_26.md and distilled into
generated/summaries/wiki_development.md.
Append new captures below.

2026-06-23 — Subsystem axis (com/base/ee) + cross-wiki evidence bridge

• New subsystem axis. Added a list-valued typed field subsystem: [com | base | ee] — the physical
  join key tying a claim to the loop it touches, and across concept↔code↔logs. Canonical home:
  subsystem_decomposition.md. The decomposition is not
  symmetric: only com decouples (the $3\times 3$ $m\mathbf{E}$ loop); base+ee are the coupled
  $9\times 9$ circumcentroidal block $\breve{\mathbf{M}}$ — so circumcentroidal-block mechanisms tag
  [base, ee], never one alone. Backfilled 27 control/singularity pages; query with Dataview
  WHERE contains(subsystem, "com").
• Tikhonov does NOT act on the CoM loop. Tikhonov $\mathbf{\Gamma }$-regularization (and the whole
  derate cascade) acts on the [base, ee] $9\times 9$ solve, not the decoupled $3\times 3$ CoM loop.
  “Tikhonov damps the CoM” is a category error (the error_floor.md §4
  mis-attribution). Whether near-singular CoM spikes (${\tilde{\mathbf{x}}}_c$) are a separate
  forward-Euler effect is an open question — decided by the lag-1 autocorrelation of logged
  ${\dot{\tilde{\mathbf{x}}}}_c$.
• The two workspaces study the same ${\omega }_b$ artifact from opposite sides (and weren’t linked).
  Docs has the theory (omega_b_forward_euler_instability.md);
  the sims breve_controller.md has the fix — base velocity-damping is integrated implicitly
  (backward-Euler), which “cured the ripple.” Now bridged via
  sims_implementation_map.md.
• Candor — the COM spike images aren’t where expected. I opened com_smoke.png from the earliest
  (Jun04) and latest (Jun23) runs: both show smooth, spike-free CoM force and ${\omega }_b$ at
  machine-epsilon ($10^{-14}$–$10^{-15}$) — non-singular nominal smoke tests with the ripple already
  cured. No logged figure exhibiting near-singular CoM spikes was found. Need a pointer to the
  specific near-singular run to settle §4 (flagged in ONGOING). Logs reachable via the new sims_logs/
  symlink; curated in evidence_map.md.
• COM spikes located → Tikhonov moves the CoM indirectly. User pointed me at run_PM0314
  (logs_Jun21_26/checkin_jun21_26, an A/B over the Tikhonov onset). The Low Tikhonov onset 0.020
  variant rides to ${\sigma }_G\approx 0.005$ (near-singular) and shows the worst CoM error ($p_c$ to
  $0.34$ m, $v_c$ to $0.22$) and violent ${\omega }_b$ bursts ($0.08$–$0.23$ rad/s) — it was rejected.
  So lowering Tikhonov degrades the CoM through singularity proximity + base↔CoM coupling, not by
  acting on the decoupled $3\times 3$ CoM loop — this refines (does not overturn) §4. Note ${\omega }_b$
  here is a real $\sim 0.15$ rad/s signal (singularity-driven), distinct from the cured machine-ε
  forward-Euler ripple. A shared $t\approx 700$ transient (in $p_c$ and $v_c$, all variants incl.
  noise-free) still needs the lag-1 autocorrelation test to confirm/deny a forward-Euler signature.
  Exhibit: evidence_map.md.

2026-06-23 (evening) — GPU cluster, a Khalil edition catch, and a board-hygiene hook

• Khalil “3rd edition / 2002” was a ghost — the PDF is the 2nd edition (1996). The inbox
  KhalilNonlinearControl.pdf self-identifies as “Second Edition” with “thirteen chapters” (title
  page + preface); the draft bib claimed 3rd ed./2002. This matters because chapter numbers differ by
  edition: ch.3 here is Lyapunov Stability, but ch.3 of the real 3rd ed. is Fundamental Properties —
  a citation to the wrong edition lands a reader on the wrong page. Corrected to bibkey
  khalil1996nonlinear, ISBN 0-13-228024-8 (the PDF’s own copyright page OCR’d this as “0-13-P8U24-8” —
  marker turned digits into lookalike letters, a reason to anchor textbook metadata to the title page).
• The GPU cluster, VRAM-ranked. workstation (RTX 5060 Laptop, 8 GB, ready) > pc1=toni
  (GTX 1660 SUPER, 6 GB, needs marker installed) > pc2=toni-toni-laptop (GTX 1650 Mobile, ~4 GB,
  NVIDIA driver not loaded → GPU-blocked). For marker, VRAM is the binding constraint, not system
  RAM — pc2’s 62 GiB RAM can’t rescue a 4 GB GPU, and an 8 GB card can OOM (out-of-memory) on a 400-page
  span, which is the real reason to convert one chapter at a time. Full note: wiki/reference/compute_cluster.md.
• marker is NOT in a conda env on the workstation — it’s pip install --user (user-site ~/.local)
  against system Python 3.10. That answers the long-standing “which env has marker?” question, and the
  same recipe (pipeline/install_marker.sh) brings up pc1/pc2.
• The “workspace install block” is local-only. It’s two layers on this Mac — the allow-read-only.py
  Bash-classifier hook (denies chained/sh -c/non-project-python) and the macOS sandbox network allowlist
  (PyPI/GitHub/Anaconda only). Neither runs on pc1/pc2, so a pip install typed in a terminal there just
  works; the block only bites commands I issue from here. (SSH to the LAN boxes does pass the sandbox.)
• Stop-hook safety pattern. The new board-hygiene hook avoids an infinite block-loop by honoring the
  harness’s stop_hook_active flag (block at most once per stop-chain) and exiting 0 on any error — a Stop
  hook must never be able to wedge a session.
• The two wikis are joined by two keys, but the join is asymmetric. Docs wiki (152 pages: 46 sources /
  94 topics / 12 results) and the sims wiki (58 pages: 48 module pages / 10 topics) share bibkey
  (wiki ↔ .bib ↔ LaTeX — symmetric, typed, enforced) and subsystem [com·base·ee] (concept ↔ code
  ↔ evidence). The catch: subsystem is a typed, queryable field only on the Docs side; on the sims side
  the same physical key lives only in prose (breve_controller.md says “base attitude + EE wrench” but has
  no subsystem: field — just code_path, flat tags, [[wikilinks]]). So today the bridge
  (dashboards/sims_implementation_map.md) is hand-typed and goes stale, links run Docs→sims only
  (zero reverse edges), and lint_wiki.py skips the symlink trees (:139 — os.walk doesn’t follow
  them), so dead code_paths are invisible. The fix is a “keep the walls, improve the gates” roadmap:
  P1 type subsystem natively in the sims pages (one additive field) → P2 generate the bridge by
  joining subsystem+code_path (kills the staleness) → P3 reverse back-links, P4 an “airlock” index that
  routes claim→code→evidence on command, P5 an evidence-citation contract (run_ID+figure+npz), P6 extend
  lint across the boundary. Full write-up + diagrams: ../generated/wiki/connections_v1.html.