[StepenkoAnatoli]

Discussion #413: a builder can receive a snapshot that is intact but stale, and handoff OK said nothing about how old it was. This PR does three things:

• It documents what the handoff proves and what it does not (Part A).
• It makes the OK line give the snapshot's age (C1, ADR-0187, which lifts the freeze for this one item).
• It pins the whole stale-page path in an offline end-to-end test (Part B).

Commits (one task each)

1. docs: handoff OK proves the corpus is intact, not that its pages are current.
   • README builder section and the "Try it" pointer.
   • freshness-recheck SKILL.md step 2: --refresh-days 0 against --force, and what --only limits.
   • CHANGELOG.
2. handoff: the OK line says how old the snapshot is (ADR-0187).
   • verifyHandoff returns fetched: { from, to, oldestDays }, or null.
   • The OK line ends ; captures fetched <from>..<to> (oldest N days), or ; no capture date on record.
   • Each row is dated the way stale-evidence dates it: by the capture its Raw cell names, and by its Retrieved cell when it names none (freshnessDay, now exported).
   • Only current rows count: superseded rows are left out, and an unparseable day is left out.
   • The age is at least 0.
   • The age is never a finding. The exit code and the doctor line are unchanged.
   • fetched does not reach packages: the artifact copies only ok, entries and findings.
   • ADR-0187, with ARCHITECTURE, CONTEXT, README, CHANGELOG and 4 tests in the same commit.
3. test: a page that changes after collection, end to end.
   • Setup: a local page server in its own process, http-keyless, and a temp config and kit home for each machine.
   • Collector: collect, cite, PASS, brief.
   • Then the page changes.
   • Builder: handoff stays OK and dated, and preflight PASSes.
   • Collector: --refresh-days 0 collects 0.
   • --force adds E-02, and the gate FAILs with cites-superseded-row and brief-stale.
   • Moving the citation brings PASS back, and the old row and its capture are kept.
4. docs: --force also runs every search again, closed ones included (a fix to commit 1's wording, found in review).

Verification

• Rebuilt on the rename. This change was first built on cd11b49, before the Fetchproof rename (rename: the kit is Fetchproof, with a clean break (ADR-0186) #433), and rebuilt here on the renamed main: paths fetchproof/, variables FETCHPROOF_*, temp prefixes fp-. Its ADR is 0187 because the rename took 0186. Apart from the README count, every line it adds or removes is the first build's mapped through the rename's spelling (checked by diffing the two). The red-first runs and the review below were made on the first build.
• C1 was written red first: handoff 34/3 and cli 149/2 before, then 37/0 and 151/0 after.
• Part B's red check: with evidence-supersession removed from the registry, the test fails on the preflight exit.
• Each commit passed the commit gate of its own tree, node fetchproof/bin/gate.mjs --gate commit, run by hand on the staged index with FETCHPROOF_TEST_TIMEOUT=180000. Its suite rule ran the full suite on that index (ADR-0120, ADR-0183): allow in 3679, 3835, 4005 and 4004 s for commits 1 to 4. Gates 3 and 4 ran at the same time. Gate 4 ran in a scratch worktree whose HEAD tree was commit 3's, so it saw commit 4's exact diff. The hook installed on the operator's PC is still the pre-rename kit's. It does not recognise a fetchproof/ checkout and skips the suite rule (ADR-0186, Consequences), so the gate was run by hand first. Each commit then went through the installed hook. No --no-verify.
• The README test count is 2113 (main 2104, +8 in commit 2, +1 in commit 3).
• A fresh-context breaker review found no blocker. Findings F1-F8 were fixed in commits 2-4.
• Nothing touches the network beyond 127.0.0.1. There are no Firecrawl credits, no real ~/.agents, no global git config and no keys.

To flag

• The line prints on every role. It was briefed as "on a builder", but the approved plan's C1 had no role condition, and a collector checking before a push should see what the builder will see. This is a rejected alternative in ADR-0187.
• An intermittent red in the commit gate, not caused by this change. browser-transport > LIVE: a page whose resource never arrives… failed in 2 of the first build's 6 gate runs on the operator's PC. It is a guard bug: on Windows the guard waited with no limit for a killed Chromium's exit, so it overran its 70 s. It is fixed separately on branch fix/stalled-flake (ADR-0188), in its own PR. Both PRs add rows to the ADR index, ARCHITECTURE and CHANGELOG, and change the README test count. Whichever merges second rebases those, and the count is 2117 with both in.
• The Research-Kit: primary-source research for AI coding agents #413 reply is drafted but not posted; it waits for the operator's OK.

🤖 Generated with Claude Code