TRACKING: work plan from 2026-07-29 — run 2 against meta, and the block-2 judgement calls

[mmcky]

Work plan opened 2026-07-28, kept current since. Revised 2026-08-07 against main at 3dc9aeb, with every claim below re-checked rather than carried forward — the comment thread holds the running narrative, this body holds the state.

Blocks 1 and 3 are done. Block 4 — run 2 — is the front of the plan, and block 2 is the part nobody else can do. Program state is in #16; defects in #21.

Blocks are ordered by dependency, not by size. Block 3 before block 4 was the one ordering that mattered, and it has now been honoured.

Where this sits in the repo

Added 2026-08-07, because this plan is audit-shaped and had never said so. It grew out of the 2026-07-28 audit session: blocks 1 to 4 are all audit, and until this section the word "style" did not appear anywhere in it. The repo has three live threads, and they are independent — none blocks another.

Thread	Plan	State	Waiting on
audit — validation program	#12, program in #16	0.2.0 shipped; run 2 is block 4 above	You: a token scope, a plugin update, ~40 min and a deliberate interrupt
benchmark — triage-first	#4, rubric calls in #7	#33 open, CI green	Kenko: #33 review, then the #10 checklist, then the ge_arrow follow-up. #7 is the one untouched thread
qe — the style surface	#3	Entirely scaffold: 19 items, none landed	Nothing, for items 1 and 3 — see below

The validation-program gate does not reach qe, and it is worth being exact about why. The "Explicitly not doing yet" section below holds the /audit:prs, /audit:tech-debt and /audit:translations skills back until the shared method is proven. That argument is scoped to the audit family — skills that would inherit audit/references/doctrine.md, which is what "the shared method" means. A qe style skill inherits none of it: different plugin, different references, a different upstream source of truth in style-guide, and different blockers. Run 2 could succeed or fail and would say nothing about whether the style preflight engine is sound.

So qe items 1 and 3 are unblocked today — the preflight engine and the umbrella skill body. Verified 2026-08-07: qe/scripts/ holds only fetch-copilot.sh and qe/references/rules/ only its README, so nothing has silently landed; and the upstream blockers that gate items 2, 4 and 5 have not moved — project-style-guide#6 since 2026-07-21, project-style-guide#2 since 2026-06-11.

Recorded so it can be questioned rather than inferred: style is the stated flagship, the largest recurring theme across the ~630 merged lecture PRs that justify the qe plugin at all. It currently has nineteen unlanded items beside three audit releases in a week. That is not a decision this plan made — it is where the sessions went. The only real coupling between the threads is maintainer attention, which is a scheduling fact and not a dependency, so none is asserted here.

Block 1 — merge what is waiting ✅

• Reviews: run 1 of the audit validation program (action-translation, 2026-07-28) #22 — the run-1 record (reviews/audit-run-action-translation-2026-07-28.md). Merged 2026-08-03 (608992d), before audit: re-derive doctrine §4 from evidence the first run actually produced #24 as this block required, so the run-record links in audit: re-derive doctrine §4 from evidence the first run actually produced #24's documents resolved on landing.
• audit: re-derive doctrine §4 from evidence the first run actually produced #24 — doctrine §4 re-derived from measured evidence. Merged 2026-08-03 as audit 0.1.4, not 0.1.3: it was rebased onto the version guard from Release procedure: version bumps deliver, changelogs record, CI enforces #28, which caught that 0.1.3 was already published from main and would have shipped four changed files invisibly.
• Reviews: record the wald_friedman triage run (first no-candidate evaluation) #15 — the wald_friedman triage record. Merged 2026-08-03 (7d7336f) — reviews/triage-run-wald_friedman-2026-07-27.md.

This block has refilled and re-emptied twice since it was written, which is why it is worth keeping rather than deleting: #26, #27, #28 and #29 landed the same day as its own contents, then #30, #31 and #32 on 2026-08-03, then #34 and #35 on 2026-08-07. Nothing is open in this repo as of the revision except #33, which belongs to the benchmark thread below.

Block 2 — the judgement only you can supply

Unchanged since this plan was written, and the only part of it that cannot be delegated.

• Checks 9 and 10 on the run-1 bundle: does the tiering in 01-issue-triage-report.md match what .dev/PLAN.md actually says, and would you act on it? Both are still marked Maintainer's call — pending in the merged record (reviews/audit-run-action-translation-2026-07-28.md, lines 47–48), so it merged honestly as-is — append afterwards, or comment on TESTING: validation program for /audit:issues — run it before generalising the method #16.
• Decide the bundle's fate. It is in ~/work/quantecon/action-translation/.dev/scratch/audit-2026-07-28/ — matched by that repo's .gitignore rule .dev/scratch/*, so it survives but cannot be committed where it sits, and .dev/audits/ still does not exist there. Options: move it to .dev/audits/2026-07-28-issues/ now (pre-empting the convention), or leave it pending QuantEcon/QuantEcon.manual#140 — still open, last updated 2026-08-02, so waiting remains available.
• One finding in that bundle is still unactioned. feat(glossary): add Japanese (ja) translation glossary action-translation#69 has been open since June and re-creates a directory that Wave 1 deliberately deleted; still open, last updated 2026-08-01. Corrected 2026-08-07: this box originally named two findings. The other — the hold on Apply localisation rules on the sync path for newly-created files action-translation#225, measured in #227 to regress the path it touches (93.3% → 33.3%, p = 0.002) — needs nothing. That PR already carries two comments from you dated 2026-07-27, the do-not-merge hold and the confirmed regression, both of which predate the 2026-07-28 run, so the bundle's "unsent drafted comment" was already redundant when it was drafted. #225 is still open and untouched since 2026-07-27.

Block 3 — the two severity-1 defects (#21) ✅

Done — shipped as audit 0.2.0 in #34, merged 2026-08-07 (d4e8df8) and tagged audit--v0.2.0. They went in as the single doctrine-level PR this block proposed, on the grounds it gave: both concern whether an audit's own record can be trusted. #35 then brought the tutorial and CATALOG in line, since step 4 had been teaching the resume rule #34 replaced — in the very document used to test resume.

• Defect 1 — [verified] citations are not checked for reachability. doctrine §2 required file:line, a merged PR, or a tag, but never that a cited commit be an ancestor of the ref the audit names. Run 1's headline finding cited a commit that exists only on an unmerged branch. Fixed, and broader than this box described: reachability is now required of every citation form, because file:line and PR citations are ref-relative too and fail the same way; phase 2 runs git merge-base --is-ancestor <sha> <ref> before tagging a commit. Copilot's review then found a third instance of the same class — §1 rule 1 still carrying its own copy of the accepted-forms list — fixed at the cause in 27a5db6 by deleting the duplicate rather than syncing it.
• Defect 2 — the closed side is never checkpointed. findings.md held the 56 open issues; the 62 closed ones went straight to the catalog. Fixed: findings.md now carries ## Open and ## Closed, the resume rule partitions issues.json by state and resumes each side independently, and the skill says explicitly not to infer progress from a single block or from the file's length.

Block 4 — run 2, now unblocked

• gh auth refresh -s read:project first. Re-checked 2026-08-07: still outstanding. The token has gist, read:org, repo, workflow, write:packages — no read:project, which was run 1's single conclusion-changing gap.
• claude plugin update before starting. Added 2026-08-07. The installed audit@quantecon is still 0.1.4, and the plugin cache is keyed by version string, so a session that does not update executes the pre-fix procedure — the one without the reachability check, and with the half-covered checkpoint — while nothing in the output would reveal it. That would waste the most informative run in the matrix in the one way that is undetectable afterwards.
• Run 2 against QuantEcon/meta (161 open / 319 items, ~39 min estimated). Per TESTING: validation program for /audit:issues — run it before generalising the method #16's matrix it is the sharpest test of what generalises: org-wide issues with no code to verify against, which directly stresses doctrine rule 1 and may reveal that [verified] means little for a decision record.
• Interrupt it deliberately around 20–30 findings and restart (tutorial step 4, now updated to the post-audit: a citation must resolve on the ref the audit named, and both passes get checkpointed (0.2.0) #34 rule). Run 1 went straight through, so resumability — the claim this whole program exists to check — is still unvalidated. Interrupt during the closed pass if the chance comes: that half was not checkpointed at all before audit: a citation must resolve on the ref the audit named, and both passes get checkpointed (0.2.0) #34, so nothing has ever resumed from it.

Why block 3 came first, and what it bought. Run 2's headline purpose is testing resume, and defect 2 meant the checkpoint covered only half of phase 2. Testing a resume path against a checkpoint already known to be incomplete would have produced a failure that said nothing about resume logic and everything about a missing block. That is now fixed, so an interrupt tests the claim rather than a known hole.

Run 2 also produces the second data point that #23 (does the phase apparatus need to shrink?) is explicitly waiting on.

Parallel — no dependencies, any time

• TESTING: audit run 1 on action-translation — seven plugin defects, and a cost claim off by an order of magnitude #21 defects 5, 6, 7 — bundle provenance links point at a machine-local plugin cache; fetch_tracker.py preflight does not check token scopes; one location over-claimed in a summary. All small and independent, all still untouched. Defect 3 remains blocked on QuantEcon/QuantEcon.manual#140; defect 4 was closed by audit: re-derive doctrine §4 from evidence the first run actually produced #24.
• The benchmark thread — PLAN: evaluation rubric v2 — enforce couplings, verdict vocabulary, instrument fixes (from the three-way design review) #7 is the remainder. Rewritten 2026-08-07; this box previously said the thread was untouched, which has not been true since 2026-08-06. TESTING: review + expert-parallel protocol for the benchmark skill (@xuanguang-li) #10 was answered and Kenko's two misreads filed as TESTING: benchmark triage run on wald_friedman — six plugin gaps, and a floor that measures the wrong quantity #14 finding 6 — the report leads with the number where the decision belongs. benchmark: triage first — advise before code is written; every output leads with the decision (0.4.0) #33 (benchmark 0.4.0, triage-first, review mode kept) is open with CI green, awaiting his review; his queue is benchmark: triage first — advise before code is written; every output leads with the decision (0.4.0) #33, then the TESTING: review + expert-parallel protocol for the benchmark skill (@xuanguang-li) #10 checklist, then a follow-up NumPy PR on ge_arrow. [ge_arrow] Update to JAX and compare runtime lecture-python.myst#717 was closed 2026-08-07 with a formal no-conversion verdict, the program's first externally-consequential result. PLAN: evaluation rubric v2 — enforce couplings, verdict vocabulary, instrument fixes (from the three-way design review) #7 (rubric v2 design calls) has not moved since 2026-07-27 and is the one genuinely untouched thread.
• Decide on the NEXT-SESSION.md exposure. It was committed to a public branch on 2026-07-28 and force-pushed out; orphaned commit d5f3205 remains reachable, and PR audit: resolve the read-only/working-directory contradiction, and a tutorial for running one #19's timeline links to it. Decided 2026-08-07: leave it, no action — do not re-open. The call was made on the contents rather than on the description of them. The leaked object is the 2026-07-25 version of the file: no credentials, no tokens, no email addresses, no personal data, and nearly all of it public now anyway through Land the lecture evaluation system (benchmark plugin 0.3.0: rubric v2, skill wired) #5, PLAN: evaluation rubric v2 — enforce couplings, verdict vocabulary, instrument fixes (from the three-way design review) #7 and TESTING: review + expert-parallel protocol for the benchmark skill (@xuanguang-li) #10. Accepted knowingly: the SHA stays discoverable through PR audit: resolve the read-only/working-directory contradiction, and a tutorial for running one #19's force-push event, and with it a reference to a private repository's issue number.

Explicitly not doing yet

• Does the phase apparatus need to shrink? (option B, deferred to run 2) #23 — shrinking the phase apparatus. Still deferred to after run 2, by design. One measurement cannot distinguish "over-built" from "easy repo", and run 2 is the second.
• The /audit:prs, /audit:tech-debt, /audit:translations skills (PLAN: audit — from /audit:issues to the bulk-audit family #12). The whole point of the validation program is to prove the shared method before more skills are built on it — and resumability, the method's headline claim, is still the thing run 2 exists to check.

[mmcky]

Block 1 is complete. #22, #24 and #15 all merged 2026-08-03, in the order the plan required: #22 (608992d) is an ancestor of #24 (9156b2e), so the absolute run-record links in #24's documents resolved on landing. One correction to carry forward: #24 shipped as audit 0.1.4, not the 0.1.3 this block names — details on #21.

Four PRs the plan did not anticipate landed the same day, and one changes the ground under block 3: #26 (/qe:copilot-review, qe 0.1.0 → 0.2.0), #27 (CATALOG narrowed to skills that are merged and operational), #28 (per-plugin CHANGELOG.md, scripts/check-version-bump.py and a claude plugin validate --strict job, all wired into .github/workflows/validate.yml), #29 (CI actions off the Node 20 runtime, lycheeVersion pinned). After #28 the block-3 fix is a release rather than just an edit — the mechanics are on #21.

Block 1's list is non-empty again. #30 (status truth-maintenance, including the CATALOG row about run 1; qe 0.2.2, benchmark 0.3.2), #31 (docs pointers, layout trees) and #32 (strict-validate targets from the manifest) are open and unmerged as of this comment.

Block 2 is unmoved. Checks 9 and 10 are still Maintainer's call — pending in the merged record (reviews/audit-run-action-translation-2026-07-28.md, lines 47–48). The bundle is still at .dev/scratch/audit-2026-07-28/ in action-translation, matched by that repo's .gitignore rule .dev/scratch/*, and .dev/audits/ does not exist there; QuantEcon.manual#140 is still open, last updated 2026-08-02, so "leave it pending" remains available. On the two time-sensitive findings: QuantEcon/action-translation#225 is open and unmerged, last updated 2026-07-27, and already carries two comments from you dated 2026-07-27 — the do-not-merge hold, then the confirmed regression. Both predate the 2026-07-28 run, so the bundle's drafted comment was already redundant when it was drafted; read it against those two before sending anything. QuantEcon/action-translation#69 is open, last updated 2026-08-01. The NEXT-SESSION.md decision is still open and unchanged.

Blocks 3 and 4 have not started. Both severity-1 defects are unfixed on main: doctrine §2's evidence table still asks only for file:line, a merged PR or a tag, with no ancestry requirement, and merge-base --is-ancestor appears nowhere under audit/; phase 2 still writes a single findings.md block and the resume rule still reads state from it (audit/skills/issues/SKILL.md, lines 50 and 54). Defects 5, 6 and 7 are untouched too — deliverables.md carries no rule about citing a version-pinned URL, fetch_tracker.py's preflight() checks gh auth status but not scopes, and the bundle's summary still places the #106 reference in the docstring. Resumability is still unvalidated — there is no run-2 record in reviews/ — so the block-3-before-block-4 argument holds exactly as written.

Parallel items. #15 landed reviews/triage-run-wald_friedman-2026-07-27.md, but #14's five gaps are open and #14 has not moved since 2026-07-27; #7 (2026-07-27) and #10 (2026-07-24) are likewise untouched.

[mmcky]

Re-checked against the repo and every linked thread, 2026-08-07. Two changes since the comment above; everything else re-verified unchanged.

Block 1 is empty again — cleared in full this time. #30, #31 and #32 merged 2026-08-03 at ~09:35, about two and a half hours after the comment above listed them as open. d054d13 (#32) is still the tip of main, and no PRs are open today. The #22 and #15 boxes in the body are now ticked to match.

#10 is the one thread with fresh input, and it is awaiting a response. Kenko replied twice since the last comment: 2026-08-05, confirming the benchmark is our custom tool rather than a Claude built-in and explaining the marketplace-install mechanics; and 2026-08-06, reporting that he revised ge_arrow per the previous evaluation, re-ran /benchmark:review-acceleration, and the scores moved as his changes predicted — his read is that the skill is self-consistent. That is the first data point on the benchmark from someone other than its author, and it lands directly on the parallel block's benchmark line.

Blocks 2, 3 and 4 are unchanged. Checks 9 and 10 are still Maintainer's call — pending (reviews/audit-run-action-translation-2026-07-28.md, lines 47–48); the bundle still sits at .dev/scratch/audit-2026-07-28/ with its #106 summary claims as run 1 wrote them, and .dev/audits/ still does not exist; QuantEcon/QuantEcon.manual#140 is open, last updated 2026-08-02. QuantEcon/action-translation#225 is open and untouched since 2026-07-27, still carrying your two pre-run comments; QuantEcon/action-translation#69 is open, last updated 2026-08-01. Both severity-1 defects are still unfixed on main: doctrine §2's evidence table (line 22) still asks only for file:line, a merged PR or a tag, merge-base --is-ancestor appears nowhere under audit/, and phase 2 still writes the single findings.md the resume rule reads (audit/skills/issues/SKILL.md, lines 50 and 54). audit is still 0.1.4, so the block-3 fix remains a release, per the mechanics on #21. There is no run-2 record in reviews/, so resumability is still unvalidated and block-3-before-block-4 holds exactly as written.

Parallel items otherwise unmoved. Defects 5 and 6: deliverables.md still carries no version-pinned-URL rule, and fetch_tracker.py's preflight() still checks authentication but not scopes. #7 and #14 have not moved since 2026-07-27. The NEXT-SESSION.md decision is still open.

[mmcky]

Second update today — the benchmark thread became the day's main event, so the parallel block's line about it is now inverted: it was the untouched thread this morning and is the moved one tonight. The audit blocks (2, 3, 4) did not move.

#10 answered, and the two misreads became a finding. The reply to Kenko went out: the 2.3 sensitivity question is the total-vs-verdict layering working as designed (the flip's verdict did hit the floor — the number beside it is the recomputed average, which no gate ever rewrites); his quoted rubric.py turned out to be a stale pre-#5 checkout carrying the conditional x64 form that never reached main, flagged since §1 asks him to confirm exactly that line; and his ge_arrow re-run checks out against the rubric at 3.80 exactly, with one question back — what did his report's VERDICT line say, since the profile (baseline ~0.035 s, candidate 0.053× as-used) should still read no-conversion. His two comments are the same layer-collapse in opposite directions, filed as finding 6 on #14: the report leads with the number where the decision belongs.

Direction change, decided in-session: triage-first, review kept. The wanted product is "look at a lecture and advise whether a JAX upgrade is recommended" — and the record supports it: in every evaluation to date the recommendation was decided by the triage-layer instruments, never moved by the scorecard on top. #33 (benchmark 0.4.0, CI green) makes triage the front door, keeps review mode intact for the day a conversion PR exists, makes every output lead with the recommendation (landing #14 finding 6), and cites the manual's JAX style page as the canonical decision criteria — which already carries "when to use JAX / when not to" and "Converting from Numba § Decide first" (https://manual.quantecon.org/styleguide/jax.html). Copilot's review produced one valid catch — the triage paragraph re-derived weight algebra against the skill's own single-source rule — fixed, replied, resolvable. Kenko's review is requested on the PR and on #10, with his older rubric-confirmation asks re-prioritized behind it.

The §4 acceptance test lost all its targets, and how it lost them is itself a data point. On 2026-08-05 QuantEcon/lecture-python.myst#657, QuantEcon/lecture-python.myst#624 and QuantEcon/lecture-python.myst#644 all merged with their JAX content stripped — the style halves landed, the conversions did not — after QuantEcon/lecture-python.myst#654 had closed with the lecture staying in NumPy. Zero open conversion PRs remain; the only live experiment is GPU (QuantEcon/lecture-python.myst#1023, sargent_surico), outside the CPU-calibrated envelope. §4 is parked for the next real conversion PR. Remaining upstream follow-up: QuantEcon/QuantEcon.manual#104, distilling the program's measured findings (1 s floor, addressable share, readability break-even) into the manual's JAX page.

Waiting on: Kenko (#33 review, then the #10 checklist). Unmoved from the morning comment: blocks 2, 3, 4; the NEXT-SESSION.md decision; #7.

[mmcky]

Third update today, and it closes the one thing the two comments above got wrong.

QuantEcon/lecture-python.myst#717 is closed, with a full assessment (2026-08-07 04:37). That answers @jstac's 2026-08-05 hand-off — "I'll leave this one for you" — which had been sitting unanswered and was the newest item on the whole board. The decision lands where Kenko's own report pointed: no-conversion, on the measured ground that the NumPy baseline replays the lecture in 0.035 s, under the 1 s materiality floor, so at ge_arrow's 2×2/3×3 sizes there is no speedup available regardless of implementation polish. Rule jax.1, vindicated by measurement rather than assumption. @xuanguang-li has been asked to open a follow-up PR carrying the genuinely good parts of the rewrite — the immutable NamedTuple bundle, the pure one-call API, and the four real bug fixes — in plain NumPy.

The correction the comment above needs. Both my 00:54 note here and the 00:55 direction note on #4 asserted that zero open conversion PRs remained as of 2026-08-05. That was wrong when written: #717 — the ge_arrow JAX conversion, and the PR this entire evaluation program was derived from, as QuantEcon/QuantEcon.manual#104's title records — was open, and had fresh activity on 2026-08-05 from jstac, which is what made its timestamp read as settled. The three PRs the sweep did check are confirmed: QuantEcon/lecture-python.myst#657, QuantEcon/lecture-python.myst#624 and QuantEcon/lecture-python.myst#644 each merged with zero JAX content in the diff, and QuantEcon/lecture-python.myst#654 is closed. It was the conclusion drawn from them that overreached past the one PR that survived. The claim is true now — as of 04:37 today, and true because the PR that falsified it was closed, not because the sweep was right.

§4 is now definitively parked, and one route to un-parking it is foreclosed. Kenko's 2026-08-06 re-run scored a revised ge_arrow at 3.80 against the 2.85 anchor, but that revision was never pushed to #717 — its last code commits date to February — so until this morning the acceptance test's target was one git push away. Closing the PR ends that. The proposed follow-up will not restore it either, since a NumPy restructure is not a conversion. §4 waits for a genuinely new conversion PR, exactly as the 00:54 comment said, just for a different reason than it gave.

One question on #10 is now answered from the other direction. I had asked Kenko what his re-run's VERDICT line said, since a profile of baseline ~0.035 s and candidate 0.053× as-used should still read no-conversion. The closure comment states that reasoning as the decision of record, so he will read both together.

Filed: QuantEcon/lecture-python.myst#1030 — Example 4's transition matrix in ge_arrow.md has row 2 summing to 1.8, so every published output of that example is computed from a matrix that is not stochastic. Pre-existing on main, found in the closing style-guide pass, and the promised follow-up from that comment.

Waiting on Kenko is now three deep, in his stated order of attention: #33's review, then the #10 checklist, then the follow-up NumPy PR. He acknowledged the first two last night (2026-08-07 00:58 on #14, 01:00 on #33) and said he would look that evening.

Blocks 2, 3 and 4 are unmoved, re-verified against main at 04:15. d054d13 is still the tip and #33 is still the only open PR, CI green and mergeable. Checks 9 and 10 remain Maintainer's call — pending; the run-1 bundle still sits in .dev/scratch/audit-2026-07-28/ and .dev/audits/ still does not exist, with QuantEcon/QuantEcon.manual#140 open since 2026-08-02; QuantEcon/action-translation#225 is untouched since 2026-07-27 and QuantEcon/action-translation#69 since 2026-08-01. Both severity-1 defects are still unfixed: audit is still 0.1.4, doctrine §2's evidence table at line 22 still asks only for file:line, a merged PR or a tag, merge-base --is-ancestor appears nowhere under audit/, and phase 2 still writes the single findings.md the resume rule reads (audit/skills/issues/SKILL.md, lines 50 and 54). Defects 5 and 6 are untouched — no version-pinned-URL rule in deliverables.md, and fetch_tracker.py's preflight() still checks authentication but not scopes. There is no run-2 record in reviews/, so resumability is still unvalidated and block-3-before-block-4 holds as written. #7 has not moved since 2026-07-27, and the NEXT-SESSION.md decision is still open.

Next. With the benchmark thread handed to Kenko and the conversion-PR question settled, block 3 is the largest unblocked piece of work left on this plan, and still the only thing gating run 2. Under #28's release rules it is a release — audit 0.1.5 — not just an edit.

[mmcky]

Block 3 is done, and block 4 is now the front of this plan.

#34 merged 2026-08-07 as d4e8df8 — audit 0.2.0, released and tagged audit--v0.2.0, branch deleted. Both severity-1 defects fixed as one doctrine-level PR, which is the shape this plan's block 3 called for. Defect 1's fix is broader than the block described: reachability is now required of every citation form, not only commits, because file:line and PR citations are ref-relative too and fail the same way. Copilot's review then caught a third instance of the same class — doctrine §1 rule 1 still carrying its own copy of the accepted-forms list while §2's had grown — fixed at the cause in 27a5db6 so the forms are written down once. Details on #21.

#35 is open, repo-level docs only, no version bump. It exists because the tutorial was the one document that could not be left describing the pre-#34 behaviour: step 4 taught the single-block resume rule that #34 replaced, and quoted the three-form citation list without the reachability requirement. It also strengthens two review checks where run 1 showed them weak — check 3, which passed the unreachable citation, and check 5, which held in the report and broke in the checkpoint. CATALOG records the two fixes while keeping its central caveat, which is still true.

Block 4, in order, with the two steps that are easy to skip and would invalidate the run. First gh auth refresh -s read:project, since the missing scope was run 1's single conclusion-changing gap. Then claude plugin update — the install cache is keyed by version string, so a session that does not update runs the cached 0.1.4 and executes the pre-fix procedure while nothing in the output reveals it. Then /audit:issues QuantEcon/meta (161 open / 319 items, ~39 min), interrupted deliberately around 20–30 findings and restarted. Per #16's matrix meta is the sharpest test of what generalises, and the interrupt is the test the program exists for.

Block 2 is unchanged and is the only thing on this plan that nobody else can do. Checks 9 and 10 are still Maintainer's call — pending in the merged run record; the bundle still sits in action-translation's gitignored .dev/scratch/audit-2026-07-28/ with .dev/audits/ still absent and QuantEcon/QuantEcon.manual#140 still open; QuantEcon/action-translation#69 is the one genuinely unactioned finding, since the #225 hold was already recorded on 2026-07-27 before the run drafted it.

Parallel items. #21 defects 5, 6 and 7 remain untouched and independent; defect 3 still waits on QuantEcon/QuantEcon.manual#140; defect 4 was closed by #24. On the benchmark side, #33 is open with CI green and Kenko reviewing, and his queue is #33, then the #10 checklist, then the ge_arrow follow-up PR. #7 has not moved since 2026-07-27.

The NEXT-SESSION.md exposure line on this plan is closed — decided, no action. The decision was made on the contents rather than on the description of them: the leaked object is the 2026-07-25 version of the file, and it carries no credentials, no tokens, no email addresses and no personal data. Nearly all of it is public now anyway through #5, #7 and #10, and the appendix is a collaboration email whose substance was subsequently posted publicly. Accepted knowingly: the SHA stays discoverable through PR #19's force-push event, and with it a reference to a private repository's issue number. This item has resurfaced in every status pass since 2026-07-28 and should not be re-opened.

[mmcky]

Correction to the comment above: #35 is merged, as 3dc9aeb at 2026-08-07 05:22 — it was open at the time of writing and landed shortly after. Repo-level files only, so no version moved and no tag was owed. Copilot's round on it is closed: three comments, all valid, all replied. Two are worth carrying forward as method rather than as edits. The tutorial's absolute blob/main link to doctrine §2 was inconsistent with line 7 of the same file and with developing-skills, which says relative links stay relative because CI checks each one resolves — the AGENTS.md absolute-URL rule governs links pointing out of a plugin directory, since an installed plugin ships alone and cannot resolve ../, and files under docs/ never ship alone. And review check 3 had become the fourth copy of the citation-forms list, so it now enumerates nothing at all rather than being synced — the same fix applied to doctrine §1 on #34, and the reason both defects existed in the first place.

Block 4 is unchanged and is now the front of this plan; block 2 is unchanged and is still the only part nobody else can do.