Triage 2026-07-20: priority order for the 33 open issues — silent-corruption class first

[mmcky]

Triage of all 33 open issues as of 2026-07-20, with a priority order to work from tomorrow. This supersedes the ordering in #113 (2026-07-18) — not its content, which still holds, but its sequence, because five issues opened after it (#115–#119) land squarely on the argument #94 made and change what should go first.

Why the order changed

#94's second ordering rule was detection before repair: "fixing five silent bugs without a check that would catch the sixth is worse value than building the check." That was written on 2026-07-15 against five known defects. Since then the lecture-python.zh-cn resync wave produced the sixth, seventh and eighth — #118, #119, #117 — and every one of them followed the predicted shape exactly: the run reported success, the output looked plausible, and the corruption surfaced only weeks later on a cold strict build in a downstream repo.

#118 is the sharpest case. The resync wrapped an entire 1,465-line document in a code fence and wrote V, σ, μ as its title — a Python tuple-unpacking line lifted from line 686 of the source. It presented downstream as 13 × "headings start at H2, not H1", and the fix recommended there would have restored an H1 and left the swallowed {include} broken. Nothing in our pipeline noticed; nothing in our pipeline could have noticed. That is the whole case for Phase 2 in one file.

The three fresh defects also sharpen the shape of the fix. #118 (structure parse fell through to a degraded path), #119 ({raw} jupyter → {raw}) and #65 ((label)= anchors dropped) are all one class — structural MyST mutated by a path that should have passed it through verbatim — and #119 proposes a guard that is cheap, deterministic, and catches the class: assert every directive info string in the output matches its counterpart in the source, argument included. Those lines are never legitimately translated, so any divergence is a defect by definition.

Priority table

P	Issues	Theme	Why this band
P0	#118, #119, #65, #90, #94 Phase 2	Silent corruption + the detection layer	Output is wrong and the run says success. Eight instances now, discovered downstream every time.
P1	#117, #107, #106	Resync correctness gaps	Deterministic, known-reproducible, each has evidence attached; #117 and #106 have cheap first-cut fixes.
P2	#113 item 1, #115, #116, #109, #95	Delivery and mechanical polish	Small, unblocking. The estate bump is the one that decides whether any P0 fix reaches a target repo.
P3	#82, #92, #103, #91, #66, #53, #94 Phase 6	Measurement, contracts, consolidation	Real value, no active bleeding. #82 gains its first metric free from P1.
P4	#70, #56, #55, #61, #74, #73, #81, #48, #1, #2, #3, #4, #6, #7	Features, design, long-horizon	Nothing here is blocked by anything above.

P0 — the corruption class

The four items are one piece of work, and the order matters:

1. Directive-shape assertion on write (forward resync stripped a MyST directive argument: {raw} jupyter -> {raw} (sir_model, v0.17.0) #119, Translator drops MyST label anchors like (label)= before headings #65). Cheapest, most immediate. For every fenced directive in the output, the info string must match the source's, argument included. Fail the file loudly on divergence. This is a guard we can ship before understanding forward resync wrapped an entire document in a code fence and mis-derived its title (ak_aiyagari, v0.17.0) #118, and it independently closes the Translator drops MyST label anchors like (label)= before headings #65 production breakage that has been open since May.
2. Round-trip invariant reconstruct(parse(doc)) === doc (Work plan: suggested order for the post-review backlog #94 Phase 2). The detection layer. Work plan: suggested order for the post-review backlog #94 already establishes it catches three of Silent data loss in the sync merge path — five ways translations vanish while the run reports success #90's five defects as a class, plus the typography bug that shipped. On the evidence of forward resync wrapped an entire document in a code fence and mis-derived its title (ak_aiyagari, v0.17.0) #118 it would also have caught a document that came back fence-wrapped.
3. Root-cause forward resync wrapped an entire document in a code fence and mis-derived its title (ak_aiyagari, v0.17.0) #118. Why did structure parsing fall through to a degraded path that both mis-derived the title and treated the body as opaque? The file is the largest in its edition (1,465 lines, 59 code cells) and a JAX rewrite — size, a parse timeout, or a truncated model response are the starting hypotheses. A sweep of all 8 edition clones found no second instance, so this is rare, which makes it more dangerous, not less: rare and silent means it lands unnoticed.
4. Real validateMyST + canonicalizeHeading(), then the Silent data loss in the sync merge path — five ways translations vanish while the run reports success #90 remainder. Unchanged from Work plan: suggested order for the post-review backlog #94 §3–4. validateMyST validates nothing today because parseSections never throws, so the gates cannot fire.

Also relevant to P0 and outside our repo: the downstream editions lacked -n -W on the notebook build, which is why #118 and #119 survived merge. That has been fixed in QuantEcon/lecture-python.zh-cn#185. Worth confirming the same guard exists across the other editions — it is our last line of defence and it was absent where it mattered.

P1 — resync correctness

Issue	Cheapest useful fix	Evidence
#117	Lint: extract every {cite} key from the resynced file, assert it resolves in the target .bib, pull or flag if not	4 of 36 lectures in one tier; all four keys present upstream, absent downstream
#107	Validation wave first — the prompt change is live but unmeasured	11 hand-fixed intro.zh-cn branches are the ground truth
#106	--from-status so forward can consume a check-sync run; fix summary to bucket by outcome not verdict	Discovery found 12 of 41 stale files; summary reported files as skipped that it had resynced

#107 is the one to be careful with. The prompt change shipped in v0.18.0 and nobody has measured it. #113 item 2 already specifies the wave — re-run forward -f on long_run_growth, inequality, heavy_tails, diff against the hand-fixed branches, count CJK-bearing code lines per file. Run it before layering more prompt changes on top, and it seeds #82 with its first real metric as a side effect.

P2 — delivery

The estate bump is listed first for a reason: the estate sits at v0.16.1 on 9 explicit ref: pins across 4 repos, two releases behind. Every fix in P0 and P1 lands in a repo that is not running it until those pins move. Repos on @v0 are already current.

#115 and #116 are both small and both bite during exactly the drift-recovery waves that are surfacing everything else — #115 leaves 60+ resync PRs with no rebase trigger because the template filter only matches translation-sync-* and forward writes resync/*; #116 is a one-line trailing-newline fix that dirtied all 69 branches in the Track B wave. #109 carries one decision worth making deliberately rather than by default: whether createForwardPR should also apply action-translation so review triggers on resync PRs at all. #113's recommendation is yes.

Suggested order for tomorrow

Start at P0 item 1 — the directive-shape assertion is a contained, shippable guard that closes #119 and #65 together and gives immediate protection while the larger Phase 2 work proceeds. Then the #117 cite lint, which is the same shape of cheap deterministic check on a different surface. Both are small enough to land in one release with the P2 mechanical items (#115, #116), which would make the estate bump worth doing once, afterwards, carrying everything.

The #107 validation wave can run in parallel with any of the above — it is measurement, not code, and it does not conflict.

Gardening

• Design: should French typography apply to the action's sync path, and how? #81 — likely closeable; v0.17.0 shipped typography in the sync path (fix(sync): apply typography in the sync path, typeset heading map with body #99). Verify and close.
• Work plan: suggested order for the post-review backlog #94 Phase 8 lists ~10 stale issues with close rationales already written.
• node24: GitHub is already force-running the action on Node 24 (+ the @actions/* majors that clear the last advisories) #89 is mostly done (chore(runtime): Node 24 declarations + @actions/* CJS majors — 0 prod advisories #111); what remains is the ESM-only @actions/* 3.x/9.x migration, which needs a Jest ESM story and has no deadline now that advisories are at 0.

Refs

#113 (post-v0.18.0 plan, superseded in sequence only), #94 (standing work plan, still the reasoning of record), #114 (2026-07-18 session handoff).

[mmcky]

2026-07-21 status against this triage — P2's estate bump is done (out of order, deliberately), plus two releases the triage didn't anticipate.

The one sequencing change from working this list: the P2 estate bump went first, ahead of P0. Reasoning recorded in project-translation's PLAN: every P0 fix lands in a repo that is not running it until the pins move, and moving them to floating @v0 is a one-time cost that makes every subsequent release self-delivering. That bet paid out twice within the day — v0.18.1 and v0.19.0 both reached the whole estate on tag-move with zero rollout PRs.

Band	Movement
P2	✅ Estate bump complete — 9/9 programming-trio pins on @v0, verified live; project-translation#9 closed. ✅ #115 fixed (v0.18.1) — and rescoped: the prefix was enforced in three places, not the template alone; src/branch-naming.ts now owns the prefixes with structural drift tests. #116 still open. #109 partially done: harness targets aligned to production shape (review + rebase on @v0, PAT steady-state) and now load-bearing — v0.19.0 was validated on the harness before release.
P0-adjacent	The -n -W question this triage asked ("worth confirming the same guard exists across the other editions") is answered: it did not — .fa and .fr both built non-strict. Now closed across all five editions, and .fr's first strict run caught three silently-broken {include}s on the spot (.fr#14). #117 gained a second instance class (shared _admonition/ assets; presence ≠ translated, see fa#138).
New since triage	#123 (rebase's overlap gate never fires between wave PRs — found by harness test, fixed by #124/v0.19.0 as opt-in rebase-stale-siblings) and #125 (open decision: GITHUB_TOKEN-pushed rebase heads get zero workflow runs, so conflict-rebased content currently lands unreviewed in production; PAT rollout to the five edition repos awaits sign-off).
P0 proper	Unstarted, and next. Order stands as written: directive-shape assertion (#119, #65) → round-trip invariant (#94 Phase 2) → root-cause #118 → real validateMyST. One input to that work from this week: src/index.ts cannot be unit-tested at all (import.meta.url vs Jest's CJS registry) — that structural gap let the #115 fix miss its third layer, so Stage 2's guards must land in testable modules, not index.ts.
P1	Untouched (#107 validation wave still the concurrent-anytime item).

Working plan of record for the sequencing: project-translation PLAN.md.

[mmcky]

Triage refresh 2026-07-22 — 31 issues open. The band structure of this triage still holds; three days of work have hollowed out its top band, and one new production bug jumped the queue.

Movement since the 2026-07-21 status comment

P0 item 1 is done: v0.20.0 shipped the structural parity guard on every write path (#128), closing #119 and #65, and the @v0 estate picked it up on tag-move. #125's PAT rollout completed and closed the same day. New since then: #131 — forward --github PR labels never match the review workflow's filter, so every CLI resync PR skips AI review silently (all six mini-wave PRs did). That is the last live instance of the "safety net silently absent" class, so it went to the top: fix is in #132 (option a — the CLI now applies action-translation alongside its existing labels, keeping the template's label contract single-valued), which also carries the resync-drifted tutorial fix (SOURCE_ONLY → init -f, not forward -f).

Updated priority order

P	Issues	Status vs the 2026-07-20 table
P0	#131 (PR #132), #94 Phase 2, #90	Round-trip invariant + real validateMyST + the #90 remainder are now the whole band. #118 drops out (below).
P1	#118 root-cause, #117, #107, #106	#118 downgrades from P0: the parity guard means a recurrence fails loudly at write time, so what remains is investigation, not an open wound.
P2	#116, #109 remainder, #95, #92	#109's target-side work is done and battle-tested; only the source harness repo remains (ref: v0.16.1, now four releases stale).
P3	#103, #82, #66, #91, #53, #74	#103 is newly unblocked: with #125 done and #131 fixed, review coverage becomes complete — its precondition. #74 needs exactly two things (myst#979 + one e2e sync), tracked live in project-translation#14.
P4	#70, #56, #55, #1, #2	Unchanged.

P0 notes

One decision now gates the #90 work and should be made first: defect 2 (target-only sections silently deleted) conflicts with the v0.20.0 parity guard — the naive fix would trade silent deletion for loud false alarms (ifp_egm is a known case from the guard's corpus dry-run). The declared-additions co-design recorded in #90's latest comment is the entry point. Defect 1 is substantially addressed by v0.17.0; what remains is a verification pass that canonicalizeHeading() covers every write and lookup site. Defects 3–5 are untouched. The constraint from the #115 post-mortem stands: guards land in modules Jest can load, not index.ts.

P1 notes

#117 should be retitled/widened to "shared assets referenced by lectures are not synced" per its own comment — two member classes now (quant-econ.bib, _admonition/*.md), and the .fr seeding evidence says the check wants to run over the full target tree, not only touched files; the cite lint stays the cheapest first cut. #107's validation wave remains the run-anytime measurement item and is getting overdue — the localisation prompt change has now been live through three releases unmeasured, and it seeds #82's first metric for free. #106 wants a re-scope pass against v0.18.0's flag-based discovery before work starts; what #113 listed as remaining is --from-status and nothing else confirmed.

Gardening — likely closeable, worth an hour

• Switch translation PR identity to quantecon-services machine user #61 — partially overtaken: the Rebase-pushed commits get no CI runs under the default GITHUB_TOKEN — rebased PRs end up with run-less heads #125 rollout put QUANTECON_SERVICES_PAT into every rebase workflow. Verify whether the sync PR path is covered too, then close or narrow.
• FEAT: GPT5 reviewer of the PR -- add comment to resulting PR. #2 — review mode has existed for many releases; close as done/superseded.
• Design: Consider _translation.yml for heading-map storage #3 — superseded by the translation: frontmatter block shipped via Refactor heading-map → translation frontmatter with explicit title field #51; close.
• Add test case for MyST target anchors before title heading #6 — the parity guard covers (label)= anchors; verify its test corpus includes the anchor-before-title shape, then close.
• test: merge order for lecture-python-programming translation PRs (v0.12.3 validation) #48, docs: Update presentation for action-translation rename #7, Evaluation Tool Improvements - Human Reviewer Findings (Dec 2025) #4 — stale; Work plan: suggested order for the post-review backlog #94 Phase 8 has pre-written close rationales in PLAN.md for most of these.
• .dev/ notes convention — pilot adoption #73 — the .dev convention is adopted and in daily use; close or convert to docs.

Single most valuable next actions after #132 merges: the #90 defect-2 / parity-guard co-design decision, then the round-trip invariant — the detection layer that has been "next" since #94 was written, and the only unstarted item that catches the seventh member of the corruption class before a downstream strict build does.

[mmcky]

Final status 2026-07-22 (end of session) — this triage is executed; closing. The refresh comment above is superseded by the day's own movement; the standing plan of record is #94.

Band	Outcome
P0 #131	✅ Fixed (#132), shipped in v0.21.0, estate on @v0 delivered on tag-move
P0 #90 defect 2	✅ Resolved as removal-with-visibility (#137) on the 211-pair estate-scan evidence; translation.additions shelved with decision record on #90; parity guard confirmed unchanged. The P0 co-design gate is dissolved.
P1 #107	✅ Closed. Validation wave run and posted (enumerated classes hold byte-for-byte); residual class (target-local data sources) + a newly found preamble-leak defect both fixed deterministically in #140 (pin + verify + fail loudly)
P1 #118	Root-cause investigation still open, still P1 — recurrence now fails loudly at write time
P2/P3	#116, #109 remainder, #95, #92, #103, #82 (now seeded — see below), #66, #91, #53 unchanged; #134 newly filed (init -f repo-scoped side effects)
Gardening	Unstarted: #61, #2, #3, #6, #48, #7, #4, #73 — rationales in PLAN.md Phase 8

What's next is #94 Phase 2 (round-trip invariant + real validateMyST), which today handed three fresh exhibits: the model-preamble leak (invisible to the parity guard — no directives or anchors in prose), the guard's structural-tokens-only scope generally, and the fence-aware walk now existing in four copies across modules (structural-parity, target-local-reads, forward's detector, and the CLI parsers) that Phase 2's shared utilities should consolidate.

Refs: v0.21.0 release notes, #90 decision record, #107 validation comment, #140.