fix(verify): stop counting source-disclaimer sentences as citations

[askalf]

fix(verify): stop counting source-disclaimer sentences as citations

Closes #186.

When the synthesizer is transparent about which sources it discarded —

"note that sources [3], [5], [7], [8], [9], [11], and [12] contain no
information about ternary LLM accuracy and are not cited below."

— extractCiteIds pulled seven ids out of that aside, each was recall-checked
against the disclaimer sentence, each correctly failed (they ARE irrelevant),
and all seven landed in unsupportedIds. The answer was penalised precisely for
disclosing what it rejected. A live canary run scored supportRatio 0.52 with 7
of its 10 "unsupported citations" coming from that single sentence.

The fix: exclude a sentence from citation accounting when BOTH

1. isSourceDisclaimer() — it contains a source noun (sources/references/
   citations) AND an exclusion phrase (no information / not cited /
   irrelevant / does not address / …), and
2. EVERY cited id in it also fails the recall threshold.

Requiring both is the safety property, and it is the whole design. Condition 2
means a sentence with even one supported cite is a real claim and is never
skipped, so this cannot quietly bury a partially-hallucinated citation.
Condition 1 is deliberately narrow — an exclusion CLAIM, not any negation — so
a genuine negative finding like "the trial showed no improvement [4]" is
untouched, having no source noun.

Excluded sentences are RECORDED in report.disclaimers rather than dropped, and
the CLI prints a line when any are excluded. An exclusion nobody can see is
indistinguishable from a miscount, and the point of this change is to make the
score honest, not quieter.

Residual gap, stated plainly: a fully-unsupported claim that also names
"sources" and negates them would be skipped. That phrasing is rare, and it
still surfaces in disclaimers rather than vanishing.

Tests (+6 in test/verify.test.mjs, 33 in that file, 859 suite-wide):

• the disclaimer is excluded AND recorded
• a disclaimer-shaped sentence with ANY supported cite still counts
• a fully-unsupported ordinary claim is still flagged (the fix must not
  weaken the hallucinated-citation detection it sits beside)
• isSourceDisclaimer needs both signals; "showed no improvement [4]" and
  "does not address long-context recall [4]" are NOT disclaimers
• end-to-end: the issue's repro goes 0.13 -> 1.00

Verified the new tests FAIL with the skip disabled (2 fail) and pass with it,
so they are regression guards rather than tautologies. Report shape change is
additive; the only consumers (agent.ts, cli.ts) read .unsupported.

[sprayberry-reviewer]

This is an automated review from the Sprayberry Labs fleet code reviewer.

Verdict: No blocking issues found — approving.

What I checked

• gh pr diff 191 (src/cli.ts, src/verify.ts, test/verify.test.mjs)
• CI rollup: gh pr checks 191 — CodeQL, actionlint, analyze, build(20), build(22) all pass
• Traced the counting-loop refactor in src/verify.ts line-by-line to confirm the new two-phase
  (score-then-tally) approach is arithmetically equivalent to the old per-id inline increment for
  every non-disclaimer sentence: totalCitations += citedIds.length; supportedCitations += citedIds.length - unsupportedIds.length matches the sum of the old per-iteration += 1 / += 1.
• Confirmed the disclaimer-skip gate (unsupportedIds.length === citedIds.length && isSourceDisclaimer(sentence)) requires every cite in the sentence to fail recall, so a sentence
  with one genuine supported cite is never dropped — this is the safety property the PR claims, and
  test/verify.test.mjs exercises it directly ("a disclaimer-shaped sentence with ANY supported cite still counts").
• Checked the only two consumers of the report shape (src/agent.ts:825, src/cli.ts) — both read
  .unsupported, which is unaffected in shape; disclaimers is a pure addition.

What's good

• The isSourceDisclaimer regex pair (source-noun + exclusion-phrase, both required) is deliberately
  narrow, and the test suite specifically guards the false-positive case ("The trial showed no
  improvement over baseline [4]" / "does not address long-context recall [4]") that would otherwise
  swallow a real negative finding.
• Excluded sentences are recorded in report.disclaimers and surfaced in the CLI output
  (src/cli.ts:615-619,633) rather than silently dropped — auditable, not just quieter.
• New tests were verified to fail with the skip disabled per the PR description, so they're regression
  guards rather than tautologies.

Minor (non-blocking): r.disclaimers?.length in src/cli.ts:617 uses optional chaining even though
disclaimers is a non-optional field on VerificationReport — harmless defensiveness, not worth a
changes-requested.