Work plan — 2026-08-10 backlog review: all 67 open issues triaged, fix-status verified against v0.25.0 (supersedes #94, #198)

[mmcky]

Tracking issue for the 2026-08-10 review of all 67 open issues, verified against main @ v0.25.0 (3db2f9f). Every issue was read in full; bug claims were checked against the current source, CHANGELOG.md, and comment threads (several issues carry status-changing comments their bodies don't reflect). Labels follow QEP-0002 (software set). This is a tracking parent and carries no Type label; the work phases are sub-issues.

Where we stand (verified 2026-09-01 19:57 AEST)

Fact	Measured 2026-09-01
main	76d36d7 (#299) — suite 1,538 green, lint --max-warnings 0, 0 prod advisories
Latest release	v0.27.0 cut 2026-09-01 06:45Z (#298); v0 = v0.27 = v0.27.0 = 125801c — alias moved, no #109-class lag. Headline: ml native-review round 2 (#297, glossary v0.4.0, rules 18 → 24, first deterministic ml lints), prompt caching on every translator call (#293), tool-review-injection/ (#285), runReview extraction (#278, #169 slice 1)
W1 (#259)	not started — 0 of 7 boxes; retargeted v0.28.0 (v0.27.0 shipped without it). #169 slice 1 landed (#278); slice 2 (src/github-content.ts) is next
Open contributor PRs	#289 (#156 fix) and #291 (#254 interim) — both reviewed and upgraded in place 2026-09-01, CI green, mergeable, awaiting kp992's look. Merge order #289 → #291; src/ composes cleanly in either order (#291 moved to src/toc-captions.ts), remaining cross-conflict is the CHANGELOG Unreleased header + bundle map
#225 (sync localisation)	closed 2026-09-01 unmerged — three stacked holds (#118 cause unfixed, D2 removes the path, D1 gates the need); #178 stays open
#69 (ja glossary)	open, conflicting, last commit 2026-06-25; four reviewers' rulings (names stay English, if-in-doubt English, lake モデル, Wikipedia cross-check) not yet applied; the promised lessons-learnt reply is outstanding since 2026-08-17. #300 filed for the Wikipedia cross-check idea
Resync moratorium	in force until #276's guard ships (W1 box 7); #281 records that the failure-issue template still prescribes \translate-resync
Open issues	79 (67 at review) — twelve post-review arrivals not yet triaged into phases: #280 #282 #283 #284 #286 #287 #288 #290 #295 #296 #300 (+ #271 discuss, unplaced); see the arrivals section
Shadow-freeze lift	~2026-09-01 has arrived — W4 (#262) prompt work is now unblocked in principle; any model change still gates on #82's eval set
Tracker type	Project (set 2026-09-01 on the owner's instruction; was Task). Not yet registered in the status-projects dashboard's projects.yml — a registration PR against QuantEcon/status-projects is the remaining step for it to be observed

Resume pointer (next session): (1) merge #289 then #291 once kp992 has looked — or after a week, since both are maintainer-upgraded and CI-green; (2) triage the twelve arrivals into W1–W6 (at least #282, #287, #290 look W1-shaped — no run / state advanced past unapplied changes / state-file writer clobbers keys); (3) #169 slice 2, then W1 proper; (4) #69 — apply the thread's rulings and post the reply.

The finding

The backlog has one dominant failure shape, named by #90 and re-confirmed by every production incident since: the failure path produces a success-shaped artifact. Files declared in sync metadata but never delivered (#90.3 / #156 / #222 / #256 — five field instances, the latest on v0.25.0, 2026-08-10); unbuildable PRs shipped after the run logged its own errors (#156); localisations silently overwritten and served to readers for 17 days (#254) or as tofu for months (#182); runs that never start and never say so (#234); and — arriving after this review — resync regenerating from the source PR's merge-time snapshot so that newer source edits are silently reverted and the PR merges clean (#276, the class's sixth field instance, 2026-08-18).

The ordering keeps the doctrine #94 validated: decisions and external clocks first; detection before repair; foundations before dependents.

Verified status changes (issues whose bodies are stale)

Issue	Verified against v0.25.0
#89	Done in substance — action.yml node24, esbuild target node24, engines >=24, @actions/core 2.0.3 + @actions/github 8.0.1, npm audit --omit=dev = 0. ESM-only 3.x/9.x residual = #177 F35's trigger. → close
#90	Defect 1 substantially fixed (v0.17.0; verification pass left), defect 2 resolved (#137). Defects 3–5 live as a class, with a correction: the four index.ts fetch sites are already fail-closed (:525 collects into fetchErrors, :533–535 throws before any branch reset; :1097/:1158/:1190 → result.errors → run fails + failure issue). The live gaps are (a) drops that record no error at all — the fresh 2026-08-10 repro (lecture-python-programming.fr#29) reported success — and (b) a partial PR still ships when the error set is non-empty (index.ts:902)
#256	Defect 3 (.bib) verified fixed 2026-08-11 on the harness against the deployed @v0=v0.25.0 (new scenario 27; evidence in #256 — the target PR delivered the backfilled entry). Defects 1/2/4/5/6 live. New finding for W1: metadata files[] under-declares (bib + state delivered, undeclared)
#134	writeConfig half fixed (#246); the Phase-4 asset clobber is live (init.ts:495 unconditional copyNonMarkdownFiles)
#118	Forward now runs checkStructuralParity post-finalize (forward.ts:384) — the 2026-07 fence-wrap would be refused; residual = title-sanity check
#253	Confirmed: grep -c findTargetLocalReads dist-action/index.js = 0; guard wired into forward only
#230	Fixed and closed 2026-08-11 (#268): PRICING gains opus-5/fable-5/haiku-4-5, costUSD warns once per unpriced model, VALID_MODEL_PATTERNS gains opus-5/fable-5

The phases (sub-issues)

Phase	Priority	Gate
W0 — decisions, closures, S-size fixes	CLOSED 2026-08-11 (#258 — all twelve boxes; PRs #266–#269)	none
W1 — make partial delivery loud (now v0.28.0)	P0	land #169 first/alongside (index.ts at 0% coverage; slice 1 done #278)
W2 — localisation becomes deterministic	P1	D1 font-mechanism decision; #172 choke point
W3 — reviewer, deterministic half	P1	parallel with W2
W4 — reviewer, prompt batch	—	shadow-freeze lift (~2026-09-01 — arrived; still gated on #82's eval set for any model change)
W5 — CLI hardening	P2	D2
W6 — scheduled debt (Wave 2, eval set, cleanup)	P3	capacity

W1 was filed as v0.26.0; that number shipped 2026-08-17 as the ml round-1 release (#272–#274), and v0.27.0 shipped 2026-09-01 as the ml round-2 + prompt-caching release (#297, #293) with W1 still unstarted — so W1's work now targets v0.28.0.

Dependency spine

D1 (font mechanism) ──► W2 (scope: #178/#141 live or die here)
D2 (sync first-translate?) ──► W1 wording + W5 TOC insertion
D3 (bot identity) ──► #61 execution; #221 closes
#169 ──► W1 (test the code W1 edits)
#172 ──► W2 transforms (one home, not five sites)
#171 ──► #90.1/.4 closure
W1 (error set) ──► W3 reachability + review input
shadow window ──► W4
#82 baseline ──► any future DEFAULT_CLAUDE_MODEL change

Post-review arrivals

Arrived 2026-08-19 → 2026-09-01, not yet triaged into phases (filed same-day as their own issues per the convention; placing them is the next session's first triage job): #280 rebase force-push auto-closes open sync PRs · #282 (bug) a merged source PR can produce no sync run at all · #283 identifyChangedSections heading-level collision · #284 evaluateDiff hard-fails on ~2% unparseable responses · #286 audit nomination pass · #287 sync advances .translate/state past changes it never applied · #288 forward -f duplicated frontmatter · #290 state-file writer must preserve unknown keys · #295 resyncs preferentially consume human-edited prose (three-edition evidence) · #296 ml round-2 tracker (shipped in v0.27.0; #189 Phase 3 graduation question open) · #300 Wikipedia cross-check for glossary drafting. Also filed 2026-08-27 and already closed: #292 prompt caching → #293.

Contributor PRs 2026-09-01. #289 (#156) and #291 (#254) both reviewed in full and upgraded on the contributor's branches (maintainerCanModify): byte-preserving TOC edits instead of yaml.dump round-trips (401 and 201 changed lines on the real lecture-python TOC, respectively, before the fix), #289 gains fetch-failure coverage and an accurate PR-body notice, #291 gains overlap-based part matching (identity matching lost the caption of exactly the part a lecture-adding sync touches), the rebase-replay site and 404-only target fetches. Both pre-empt W1 boxes 3 and 4 in part; the structured TOC merge (captions never taken from source) and the declared-vs-delivered assertion remain W1's. #225 closed unmerged (see close-candidates).

Triaged 2026-08-19

• resync regenerates from the source PR's merge-time snapshot — stale content reverts newer target edits #276 (bug, high-priority) — the class's sixth field instance, on the resync path: \translate-resync regenerates from the source PR's merge-time snapshot, so source edits that landed after that PR are silently reverted and the PR merges clean. On fa the stale regeneration merged after the repoint it reverted (2026-08-18, 03:37 → 03:52). Content was hand-restored the same day; a full post-incident measurement (2026-08-19: recorded source-sha vs the source's true last-touch commit, all 21 state files across the three targets, plus content probes — the ledger is in resync regenerates from the source PR's merge-time snapshot — stale content reverts newer target edits #276) found the real residue smaller and different than first thought: zh-cn is fully consistent (its original batch PR had merged; the stale regeneration was correctly discarded); fr's state is correct but numpy.md is missing #595's two edits — the regeneration silently omitted the file after the original PR was closed unmerged, a per-file silent drop that is itself W1 evidence; fa has two stale state files (pandas, polars — the regeneration overwrote the repoint sync's correct SHAs) plus a three-line pandas_panel.md divergence introduced by the hand-restore, which restored a URL form the source itself had moved off. The two deterministic repairs merged and were byte-verified 2026-08-19 (fa#158, fr#38 in the target repos — no re-translation; forward was deliberately not used on fr numpy.md, a natively-reviewed lecture); the recorded-vs-actual mismatch set is now zero. W1 gains a detection box (code-cell invariance guard); the regeneration fix itself stays resync regenerates from the source PR's merge-time snapshot — stale content reverts newer target edits #276. Until that lands, \translate-resync should not be used.
• The sync is .md-only, so hand-localised _static/ assets can never be created, updated or repaired #271 (discuss) — the .md-only sync leaves hand-localised _static/ assets (figure notebooks, rendered PNGs, data builders) outside the sync in every direction: they can never be created, updated, or repaired by it. Class analysis, not a defect in shipped behaviour; unplaced pending a scoping decision. The single .ml orphan instance is recorded and deferred on the workspace side (QuantEcon/workspace-lectures#48, decision 2).

Close-candidates

• Apply localisation rules on the sync path for newly-created files #225 — closed 2026-09-01 unmerged: the localisation-rule prompt measurably induces forward resync wrapped an entire document in a code fence and mis-derived its title (ak_aiyagari, v0.17.0) #118 (33% vs 93% structural pass, p = 0.002) and is unchanged; D2 removes sync's first-translation path; D1 gates Sync applies no localisation rules, so every organically-synced lecture lands with English figure labels and no font config #178 in W2. Reusable pieces named in its closing comment; Sync applies no localisation rules, so every organically-synced lecture lands with English figure labels and no font config #178 stays open.
• node24: GitHub is already force-running the action on Node 24 (+ the @actions/* majors that clear the last advisories) #89, Work plan: suggested order for the post-review backlog #94, Work plan — week of 2026-07-27: v0.24.0 release decision, then Phase 2 (round-trip reformulation first) #198 — closed 2026-08-10 · Moving the sync bot to a GitHub App will silently break \translate-resync — Apps always report author_association=NONE #221 — closed 2026-08-11 (D3 recorded: D-2026-08-11-bot-identity-machine-user.md) · ml packet question: is the recurring minor/terminology demerit genuine, or a reviewer misread of keep-English policy? #250 — answer feeds W4's Reviewer finding descriptions are unreproducible and sometimes describe compliant text — stable verdicts, wandering evidence #249 rules. Also closed via W0: forward CLI writes lecture files without a trailing newline #116, claude-opus-5 costs report as $0.000 — missing pricing entry silently hides spend #230, Deprecate and remove legacy heading-map: frontmatter format #53, docs: heading-maps.md documents a key format the action has never written #91, docs: Update presentation for action-translation rename #7

QEP-0002 label migration (repo level) — done, verified 2026-08-19

All four items below are applied on the repo (testing deleted, maintenance at #fbca04, core labels + refactor installed, relabels spot-checked: #169 refactor, #103 discuss, #89 infrastructure). Kept as the record of what was changed.

• Retire testing (#5319E7): ml_metrics pinned-term retention: homograph + light-verb false positive (returns) #240 → bug, ml script-ratio gate compares a mean against a p10-p90 band, so it cannot detect the distribution shift it was designed to catch #229 → bug, Malayalam (ml) test project: benchmark vs native-speaker reference, calibration with Adisankar, go/no-go #189 → tracker (no Type); then delete the label
• Recolour maintenance #1D76DB → #fbca04 (rename-in-place keeps history)
• Install missing core labels (discuss, do-not-merge, duplicate, wontfix, automation greys) + refactor (software extension)
• Per-issue relabels: tech-debt T1: Make the Action entry point importable #169–tech-debt T7: Give the CLI a shared spine: config precedence, one worker pool, shared state #175 → refactor; Design: route sync PRs by review verdict (auto-merge vs editor label) and make human review measurable #103/Decide whether positionCorrect should be deterministic — #148's deliberate residual, now measurable #155/fa: pin whether language/library proper nouns transliterate (پایتون) or stay Latin (Python) #205/Moving the sync bot to a GitHub App will silently break \translate-resync — Apps always report author_association=NONE #221 → discuss; node24: GitHub is already force-running the action on Node 24 (+ the @actions/* majors that clear the last advisories) #89 → infrastructure; high-priority only on the outliers (Silent data loss in the sync merge path — five ways translations vanish while the run reports success #90, sync opens a PR whose _toc.yml references a lecture the same run failed to translate #156, The #107 i18n-preservation guardrail is prompt prose and fails at scale — 21 lectures stripped, 9 now rendering tofu #182, sync silently first-translates lectures missing from the target — unreachable when it succeeds, invisible when it fails #222, Sync silently does not run when a merged source PR has an agent-authored head branch #234, The deterministic target-local-read guard is in neither the action bundle nor sync mode, and is blind to URL-form reads #253, Sync overwrites localised _toc.yml part captions with English, and nothing detects it #254, Six defects observed syncing a six-PR batch to lecture-intro.zh-cn (declared files silently dropped, toc clobbered, .bib never synced) #256)

Full report: .dev/log/2026-08-10-issue-review-workplan.md (landed via #265, merge 77f09da).

---

Generated by Claude Code

[mmcky]

Session log 2026-08-19 — tracker state at session end, so the next session can start from the issues alone.

Post-review arrivals triaged. #276 labeled bug/high-priority and recorded here as the class's sixth field instance; #271 labeled discuss and left unplaced pending a scoping decision. Both are described in the body's new "Post-review arrivals" section.

#276's fallout is fully repaired; the bug is not. A per-file measurement (recorded source-sha vs the source's true last-touch commit for all 21 state files across fa/fr/zh-cn, plus content probes) reversed three of the first-pass claims — worth noting because the corrections are the kind this tracker exists to catch: zh-cn needed nothing at all, the remediation on fa had itself introduced a divergence from source, and fr's gap turned out to be a silent per-file drop by the regeneration rather than stale state. That drop is a second mechanism and feeds W1's declared-vs-delivered assertion directly. Repairs merged and byte-verified (lecture-python-programming.fa#158, .fr#38); mismatch set is now zero. Ledger in #276.

W1 (#259) retargeted v0.27.0 — v0.26.0 shipped 2026-08-17 as the ml round-1 release — and its scope grew by two boxes: the #276 stale-resync guard (code-cell invariance, but not a naive byte-match — fa and zh-cn localize docstrings and comments inside code cells while fr keeps them English, measured on numpy.md), and the both-directions rule on the assertion, since files[] currently under-declares as well as over-declares.

#169 (the W1 gate) is underway: slice 1 is open as #278 — runReview extracted with 17 tests, the first ever to reach entry-point logic; CI green, awaiting merge. The issue body's figures were re-verified against main first and corrected (the file has grown to 1,579 lines; two audit claims amended). Next is slice 2, src/github-content.ts.

QEP-0002 label migration verified applied on the repo and marked done in the body.

[mmcky]

Session log 2026-09-01 (19:57 AEST) — tracker state at session end. The body now carries the contract-form ## Where we stand (verified …) heading with a live-state table; previously the only stamp was a verified against v0.25.0 in prose, which stamps nothing for the dashboard.

Landed since the 2026-08-19 stamp (other sessions' work, recorded here for continuity): v0.27.0 released 2026-09-01 (#298; v0 alias moved to 125801c, verified) carrying ml native-review round 2 (#297 — glossary v0.4.0, rules 18 → 24, first deterministic ml lints, decision record D-2026-09-01-ml-exercise-statements-stay-english), prompt caching on every translator call (#293 — telemetry semantics changed: input-tokens is now the uncached remainder), tool-review-injection/ (#285), and the runReview extraction (#278, #169 slice 1). W1 did not make v0.27.0 and retargets v0.28.0 — 0 of its 7 boxes are ticked.

This session reviewed the six open PRs and set a merge order. #294 and #297 merged and were released. #289 (#156) and #291 (#254, interim) were reviewed in full and upgraded in place on kp992's branches — both had the same defect class, a yaml.dump round-trip that turned a one-line TOC change into a whole-file rewrite (401 and 201 changed lines on the real lecture-python TOC); both now edit the TOC text in place and verify by parse. #291 also had a blocker: identity matching of parts by file set lost the caption of exactly the part a lecture-adding sync touches, which is the #254 field case; it now matches by membership overlap, with no positional guessing. Both CI-green and mergeable; merge order #289 → #291; src/ composes in either order because #291's code moved to src/toc-captions.ts. #225 closed unmerged on three stacked holds (recorded in the body's close-candidates and on #178).

Decided: #291 does not close #254 — its Fixes line was changed to a plain reference, per the 2026-08-21 design that makes W1's structured merge the end state. No positional fallback in the caption matcher (English is the honest state the coming deterministic check gates on).

Inverted premise: none in the body. One aged: the W4 gate ("shadow-freeze lift ~2026-09-01") has arrived; the model-change gate on #82 still stands.

Not done, deliberately: the twelve post-review arrivals (#280–#300) were listed, not triaged — placing them is a judgement pass that deserves its own session start, and the body names the three that look W1-shaped. #69 untouched. The tracker's issue type is Task; the convention wants Project — left for the owner.

Resume pointer is in the body under the stamp: merge #289 → #291, triage the arrivals, #169 slice 2, then W1.