test(release): realign hard-lease inventory and pin the catalog build bound in #11759 test

[diegosouzapw]

Two base-reds on the pure release/v3.8.52 tip, reproduced there (not caused by any open PR):

1. hard-session-lease-bypass-inventory — inventory drifted: src/app/api/v1/web/map/route.ts (new in feat(api): add Firecrawl Map endpoint with account fallback #15703) calls getProviderCredentialsWithQuotaPreflight exactly like web/fetch → listed as credential site, class B (same as web/fetch). src/lib/db/providers.ts gained a 4th getProviderConnectionById — the priority-0 move-to-top edit path re-reads the row after reorderConnections (fix(db): preserve explicit provider priorities on edit #15485) — count 3 → 4, same file classification.
2. 11759-embedding-registry-width-and-type — on a loaded host the cold /v1/models build exceeds the 8s CATALOG_BUILD_TIMEOUT_MS default and the route answers 503 catalog_build_timeout (fix(models): return 503 with Retry-After when a cold catalog build exceeds its time bound #13438). The test asserts catalog contents, not the time bound, so it pins CATALOG_BUILD_TIMEOUT_MS=180000 (via ??=, an explicit env still wins). Reproduced: 0/2 at default on the tip under load, 2/2 with the bound.

Not touched: chat-noauth-model-cooldown "no server-side wait inside a combo" (wall-clock < 1500 ms) also fails on the tip under load 40–50; that assertion IS the behavior under test, so it is left as is (load flake to recheck on an idle host).

Validation: both files 5/5 pass; changelog-integrity 0.

[github-actions]

CI Coverage Report

• Coverage job: skipped
• PR test policy: success

Coverage artifact was not available for this run.