Report the capture window from feed clocks and pass channel and mode to the eval hook (#98)

[thossullivan]

Closes #98.

model-eol owns the deadline and the eval-hook contract, not migration verification. This is the thin slice from the 2026-09-09 design review, recorded in docs/DESIGN_EVAL_BOUNDARY.md.

What changes

• Capture window from feed clocks. Every retiring, retired, or scheduled finding gets an in-memory capture window: the applied clock's shutdown as the last day the old model answers, plus every other clock (publisher or distribution row) with a shutdown date after today and a status other than retired. Rows without a date are never listed. The human check line appends [still answers via aws-bedrock until 2026-10-14] when such a clock exists. A tentative or earliest floor renders as at least until.
• The check --json document does not carry the field. Default output must keep validating against the pinned 0.5.2 schemas (test/fixture/schemas-0.5.2), and the fixture repo already has five findings with alternatives, so no "emit only when informative" rule keeps that pin green. Adding capture to the 0.1 report shape is a separate contract decision, noted in the design doc. SPEC permits additive fields; the PR 91 precedent kept default output byte-compatible. Your call, not mine.
• Bot bodies. PRs and issues gain a ## Capture window section from the same computation. Distribution rows now join the group digest (order-independent) so a row change refreshes the body, and metadata records capture_expires, the first date on which the section changes, so an expired alternative refreshes the body even though the digest is date-independent. The evaluate and publish phases keep binding the same eval results manifest.
• Eval hook. MODEL_EOL_VIA (the plan's channel, empty for the publisher API) and MODEL_EOL_EVAL_MODE (evaluate; capture reserved, refused today). Documented in README, docs/DESIGN_BOT.md, and the bot-config schema.
• Code spans render literally. markdownCode entity-escaped underscores, and code spans show entities literally, so custom_hub rendered as custom_hub. It now drops backticks and line breaks and keeps everything else. Pre-existing; found by review.
• Digest ordering. sortedJsonDigest used localeCompare, which returns zero for canonically equivalent Unicode strings. It now compares code units.

Operational note

Existing bot PRs and issues will refresh once after this ships, because their recorded feed_digest no longer matches the new digest input.

Verification

• npm test green at every commit.
• Cross-repo UAT sweep: ok 861 / retired 659 / retiring 43 / scheduled 1935, identical to the baseline after PR 96; the same four repos fail closed.
• Acceptance: a repo referencing claude-sonnet-4-20250514 shows RETIRED 2026-06-15 ... [still answers via aws-bedrock until 2026-10-14] on the direct scan and RETIRES 2026-10-14 (35 days) [via aws-bedrock] under --via aws-bedrock, both from the feed.
• Codex adversarial review: pass 1 found 3, pass 2 found 2, pass 3 found 3; all eight closed with the reviewer's inputs as regression tests. Pass 4 is in flight and will be pushed as a follow-up commit if it finds anything.

Implemented by Codex (gpt-5.6-luna) from a Claude brief; Claude reverted the check-schema change, moved the capture tests to findingFromRef, wrote the design note, and closed the review findings.