feat(mcp_readability): reuse unchanged findings so counts track tool changes

[akangsha7]

Stacked on #623. This is the PR that turns carry-forward on; #622, #615 and #623 are inert without it.

Summary
The judge is a thinking model, so at temperature 0 an unchanged tool surface still yields different findings run to run and issue counts drift with no explanation. With a baseline store configured, an endpoint whose tool surface and judge inputs are unchanged reuses its previous findings verbatim and makes no model call. Numbers then move only for tools a team actually touched, and every departure carries a named reason.

An absent baseline block yields the null store, so a run config that has not opted in behaves exactly as before — including the google3-mirrored one.

Changes

1. evaluator/mcp_readability/orchestrator.py — builds the store and primes it once, on one thread, before the endpoint pool starts; the per-endpoint alternative is N round trips racing each other. A prime failure sets baseline_unavailable rather than letting every endpoint look like a first run.
2. Escape hatches — force_refresh, force_refresh_products and EVALBENCH_MCP_FORCE_REFRESH=1, for Guitar one-offs where editing the mirrored config is impractical. Without them a hallucinated P0 could only be cleared by editing the tool.
3. evaluator/mcp_readability/baseline.py — is_expired ages a baseline out, staggered per endpoint so a fleet primed on one day does not all re-judge on the same later day. The run then reports baseline_expired, so the re-judge is announced rather than an unexplained count change.
4. scorers/mcp_readability/style_readability.py — run decides before it judges and appends a scope clause in a partial run. The HTML leads with a provenance banner and per-tool markers, so "not re-judged" is distinguishable from "re-judged and scored the same". Provenance is written to its own column only, so the two copies cannot drift.
5. Second pass — a re-judged tool is described from scratch, so a baseline finding it stops reporting is ambiguous: either the change fixed it or the judge did not mention it. _reconcile_dropped sends just those findings back for a verdict rather than a rewrite, so a still-valid issue is restored with its original wording instead of reading as resolved. A failed reconciliation call leaves the findings in place rather than dropping them.
6. scorers/mcp_readability/scoring.py — EndpointContext gains baseline, defaulted last so existing positional construction keeps working.
7. datasets/mcp_readability/run_config.yaml — the baseline block, commented out, documenting the store choices and each deviation reason.
8. test/mcp_carry_forward_scorer_test.py (new) — drives the real scorer with a recording judge, the only way to assert what matters: zero model calls and byte-identical feedback when nothing changed. test/mcp_readability_test.py adds the orchestrator's side of that, running it twice against a local store, plus the force-refresh hatches; test/mcp_baseline_test.py covers is_expired and its stagger.