grader(pj-s-question): batch-form citations were wrongly rejected

[im-tyler]

Ruling: grader extraction defect (a), not designed strictness

The canary (eval-20261003-2, PASS) and the three batch runs (eval-20261003-3/4, eval-20261004-1, all FAIL "the tagline line (index.html:16) is not cited") give the same correct answer — index.html:16, <p class="tagline">Furniture made on the coast</p>. The difference is citation layout only: the canary used the inline `index.html:16` — `string` form; the batch runs put the word "line" between file and number ("in index.html line 16 — the element <exact string>") and introduced the exact string as a fenced or indented block under a colon-ended heading ("(index.html:16, full line verbatim):").

Grounds for (a):

• The task's stated demand is "Answer with file, line number and the exact current string" — the batch answers supply exactly that, adjacently bound. No format rule is violated.
• The grader's own header is "tolerant of the separators a real answer uses", and its history records two prior fixes of exactly this class: single quotes accepted after the 2026-09-24 canary failed, FINDINGS_JSON accepted after the 2026-09-22 canary ("a correct scan failed a grader that understood only one answer syntax"). This is the third instance of that defect class, found by live runs.
• The mutants corpus enforces content, not typography: every pj-s-question wrong variant targets a wrong line number, a paraphrase, a missed occurrence, an edit, or no citations at all. The batch transcripts are correct on every axis the corpus pins.
• GRADER_SENSITIVITY_2026-10-03 (PR S02: grader sensitivity matrix; tighten eight graders it showed were too weak #34) tightened graders against escaping WRONG variants; its known-limit doctrine explicitly rejects strictness that "would reject the honest behaviour".

Change

• graders/pj-s-question.mjs: new signpostCitations extraction — a file+line reference (`index.html` line 16, `index.html:16`) binds the exact string that follows on the same line after a dash/colon connector, or the single fenced/indented line directly under a colon-ended heading carrying the reference. Binding is tight on purpose: the text between number and string must contain a connector and no digits (range prose like "lines 18–20" never binds), blocks bind only directly under colon-ended headings. Verbatim containment, occurrence coverage and structural checks are unchanged — every bound triple goes through the same verification.
• Outcome→summary normalization hoisted (withOutcomeSummary) so markdown and signpost extraction read the same rendered text.
• New mutants: batch-form-reference (correct — the batch layout, must pass) and batch-form-paraphrased-tagline (same layout, paraphrased block string, must fail with "does not contain"). scripts/grader-sensitivity.test.mjs now allows more than one correct reference per scenario.
• FINDINGS gains a ninth-defect note; GRADER_SENSITIVITY_2026-10-03.md regenerated via --write.
• Supplementary regrade-batch-form.json beside each of the three failed records (graded with ef04ee2): all three PASS, no reasons. Originals untouched.

Tests

• Full matrix regenerated: 75 variants — 61/61 wrong rejected for their targeted reason, 13/13 correct references accepted, 1 known limit intact, CLEAN.
• pnpm run lint, pnpm run build, full pnpm test: 2386 tests, 0 failures (web deps installed locally for the deploy-package leg).
• Note: one earlier local pnpm test run showed the known port-8901 deploy-recovery flake plus a missing-web/node_modules failure; both environmental, clean on rerun. CI runs the full matrix (grader-sensitivity.test.mjs) as usual.