D2 hygiene pre-paid:
empirical.statistic, StatTests,
tested_by, stat_test, alpha all
0-file in corpus; window prose-toxic → floor label + column
only (never in vocabulary())
Ro5 FIND-1 (substantive): fiber kind not typeable
in base Properties kind columns — property_kind_closed
walks static {unit, law} → design D8 + follow-up ticket
specodelic-ung (mechanism: closed-set walkers'
effective set = base ∪ active-pack-fiber); until then law rows carry
alpha/window additively (floor-not-ceiling)
Tooling gotcha: installed spk is stale
(no pack discovery) — use cargo run -- lint for pack
probes
Gates: openspec strict 18/18, spk lint openspec 0/0, sync-sections,
just ci green; specodelic-aal closed (implementation gated
on user approval of the proposal)
Next: approve add-empirical-registry-pack →
implement pack artifact + probes + archive (tasks 2–5); then
specodelic-0dn bioimage pilot (all three pack deps closed). Also ready:
specodelic-rjb, specodelic-ung (now unblocked)
espectacular-ybs closed (fcfead1):
ah check binary never emitted slug-collision for deployed
specs — resolve_scope's (spec,id) BTreeMap collapse ate duplicates
before detect_slug_collisions; detection now runs on the raw discovery
list (slug_collision_findings helper). RED check-binary test → GREEN;
227/0 contracts.
v0.9.1 RELEASED (authorized): version bumps +
CHANGELOG stamped (ac9856e), tag pushed → Release workflow green, 6
platform assets + checksums, crates.io max_stable 0.9.1. GH#31 commented
with the lefthook fix chain and closed.
espectacular-pli closed — "workflows red on main"
was 8 root causes deep: ci.yml unquoted-colon step name (YAML parse
failure, zero jobs, since Oct 2); docs.yml stale book/ paths (build-dir
moved to docs/_book in c626e03); the days-old unstaged t-report-stable
row deletion was the spec-gate fix (spk graph exits 1 on pristine
checkouts — dirty worktree masked it, now committed); unpinned git-main
spk → pinned 0.4.0; Coverage job missing spk; pretender 0.7.0 parse
error on &raw identifier (renamed schema_doc) +
guidance-mode lease for the smoke step; dash-vs-bash printf-format JSON
bug in parse_shim (payload now cat'd from file); two flake classes —
ETXTBSY on exec of just-written shims (bounded retry in
runner::spawn_with_text_busy_retry) and a PATH env race between the
p_precedence aggregator and its re-run test (detect_with_path seam, no
global env mutation).
Verified: CI 37130993545 all 3 jobs green (first
time since Sept 30), docs deploy green via manual dispatch, live site
serves llm.txt + release.html v0.9.1; suite 10/10 flake-free;
clippy/fmt/spk gates clean.
Decisions: slug-collision fix scoped to deployed
specs (staged-vs-staged already surfaces as overlay-conflict); pretender
smoke in guidance mode per mu5/djw stance (lease fails closed
2027-01-01); slug-collision detection before map collapse, map semantics
unchanged.
Next: board empty — fresh ticketing round or idle
for GH activity; latent quality.rs trio-runner printf-JSON (unticketed,
safe while findings numeric); TAP_GITHUB_TOKEN still user-owed (8th
release skipped tap/scoop).
All 9 open dependabot PRs closed, board empty. gh
API can't merge workflow-file PRs (missing workflow scope)
— used fix-forward via git push (SSH) instead:
#29 jsonschema 0.18→0.58 (4083b6a): real API
migration — JSONSchema::compile →
Validator::new/options,
validate() returns single ValidationError, and
custom-runner.schema.json's absolute-URL $ref to
check-output.schema.json now registered in a
referencing::Registry (0.58 resolves external refs eagerly
over HTTP; offline-deterministic). Verified on PR branch in isolated
worktree first.
#6 toml 0.8→1.1.6 (#12/#14 clap 4.6.7,
clap_complete 4.6.11) (0d74555): zero API changes; PR-branch worktree
3/3 clean reruns (one failure = known bridge-shim flake, not toml).
Gotcha: genesis-vibes 0.12.0 still transitively pulls toml 0.8.23
alongside direct 1.1.6.
#7–#11 Actions bumps (checkout v7, artifact v7/v8,
deploy-pages v5, codecov v7) (9d92be4): diffs applied verbatim; CI green
on all.
v0.9.2 RELEASED (authorized): 4d432e8 + tag →
Release workflow green, 5 platform assets + checksums, crates.io
max_stable 0.9.2, live release.html updated via manual docs.yml dispatch
on main. CHANGELOG: Fixed ETXTBSY bounded retry in runner spawn path;
Changed toml 1.1 + clap 4.6.7. n8o sweep clean (no version literals in
src/tests). Local ah refreshed 0.9.2 (was stale 0.9.0 —
phantom version-drift risk).
Gotcha: Docs deploy on tag refs fails by design —
github-pages environment protection rules only allow main (v0.9.2 tag
run rejected, benign, same on v0.9.1). Always dispatch docs.yml on main
after tag releases.
Next: board empty; idle for GH activity or latent
quality.rs trio-runner printf-JSON cleanup; TAP_GITHUB_TOKEN still
user-owed (8th release skipped tap/scoop).
Second Rule-of-5 on the proposal itself: CORR-001
(TypeSafe 0.97) — duplicate-id parity contradiction (old path
first-wins, build can't fail); fixed delta to first-wins +
collision report. Dropped stale 38-violations figure (corpus now reports
0 — post Rev-15/orphan-labeling); scoped drift gate to
corpora carrying the format doc; empty-corpus fixture; ticket ids in
design.md. Commit 46872ce.
Epic specodelic-84i filed with
children 84i.1 (implement S1+S2), 9um (pushout), p5b (writer), hya
(views re-point); deps via bd dep add; JSONL re-exported
(bd export -o — stdout default gotcha routed to
whisper).
Task: download the conversation linked in
~/Downloads/cv/zai (single URL file,
chat.z.ai/s/065f…) into a markdown file.
How: recovered the prior method instead of
re-deriving it — turu recall had nothing on chat.z.ai, so grep'd pi
session history and found the 2026-10-03 13:33 snap recording the 7-link
batch download plus its exact pipeline (api/v1/auths/ guest
token → chats/share/<id> →
chats/<share-id>/messages/batch) and script path;
/tmp/zai_dl.py from that session still existed and was
reused verbatim after three small patches (OUT →
~/Downloads/cv, links source → the zai file
itself, filename prefix NN- → zai-).
Output:
~/Downloads/cv/zai-architectural-health-review-of-specodelic.md
(96 KB, 4/4 main-path messages, thinking collapsed into
<details>, tool-calls summarized) + matching raw
.json.
Routed the reuse detail (script only in /tmp, adapted-copy path,
output naming) to whisper specodelic main-branch notes, extending the
2026-09-28 download entry.
Next: none — one-shot task; consider persisting
/tmp/zai_dl.py into a real location if z.ai share harvests
recur.
Reviewed ~/Downloads/cv/ in full (3
files were missed on first pass): 4 independent AI design reviews of
specodelic — GPT-5.6 Luna (zip + raw chat), Grok 4.5 (zip),
GLM-on-Mistral @ d055583, GLM-on-Z.ai @ stale
ad4a2c0 — all from the same four-phase brief. Consensus:
engine core sound; weight is hand-coded wiring/parallel encodings; the 3
current-state reviews independently converge on schema-as-data = the
in-flight acset epic 84i (decision-grade external
validation).
Rule-of-5 review of the findings, grounded +
TypeSafe-verified (jev-1.13.0;
TYPESAFE_API_KEY set): 9 CRITICAL/HIGH findings — 0
fabricated, 2 measured-contradicted & dropped ("3+ parallel
lifecycle FSMs" overstated — corpus delegates; Z.ai gate-demotion
contradicts AGENTS.md 2026-10-01 adoption), 2 below gate (0.54/0.47)
resolved by direct codebase inspection. Sharpest finding:
REFERENCE_TYPING (guide.rs:49) is doc-only data;
enforcement is a separate match (graph.rs:163
typing_violation) — already closed by
add-acset-core task 2.3.
Tickets: created specodelic-3a8 (P3:
openspec/dual-format metadata out of core Spec → adapter; RED =
byte-identical spk lint openspec parity) and
specodelic-mlc (P4: consolidate outbound-leaf carve-outs,
post-84i); corroboration note on 84i; parser-unification
note on p5b. Canonical register:
.wai/resources/research/2026-10-03-external-architecture-reviews.md.
Issue-review pass on the batch: 3a8
re-blocked from 84i.1 → p5b (S1+S2 doesn't
touch spec.rs — my initial blocker was artificial);
files+base_commit anchors added; 3a8
file-header criterion; mlc meter quantified. Commits
4355cc2, 0a7ec7d, e7bb19e
(gate-authored .espectacular contract-flag fixes).
Fixed README v0.4.0 → v0.5.0 (the one live
inconsistency all four reviews flagged).
Flagged: uncommitted in-flight
justfile/tests/* (guard-espectacular recipe,
refs specodelic-c0o) — a parallel session's
o8t work; left untouched.
Routed bd gotchas (no refactor type; --metadata exclusivity;
ah-check TOML rewrites) to whisper specodelic env.md.
Next: approve add-acset-core → apply
on 84i.1; p5b then unblocks 3a8;
parallel session lands o8t/c0o wiring — coordinate before touching
justfile/tests.