fix(cache): recover invalidation after failed prime

[mastertyko]

Summary

Recover cache-invalidation delivery after a transient startup baseline-prime
failure, so the first successful background poll cannot acknowledge a peer
version without invalidating already-warm routing state.

The root cause was that an uninitialized poller treated its first successful
background read as a callback-less baseline. start() now ends baseline-only
acquisition before spawning the poll task, while an explicit prime() retry
before start() remains baseline-only.

Type of change

• fix: — bug fix (no behavior change beyond the bug)
• feat: — new user-facing feature or capability
• refactor: — internal refactor (no behavior change, no API change)
• docs: — documentation only
• chore: / ci: / build: — tooling, CI, packaging
• test: — test-only change
• Breaking change

Linked issue: None — no matching upstream issue was found in the bounded related-work search.

OpenSpec

• This PR includes / updates an OpenSpec change
• Not applicable — bug fix that matches the existing spec
• Not applicable — docs / CI / chore only
• This PR touches a codex-faithful path and preserves upstream-equivalent behavior

Change directory:
openspec/changes/recover-cache-invalidation-after-prime-failure/

Changes

• Arm conservative callback delivery synchronously when background polling
  starts without a recorded version baseline.
• Preserve callback-less semantics for an explicit prime() retry before
  background polling.
• Add regressions for callback acknowledgement, a real warm upstream-route
  cache, account-routing refresh, and HTTP bridge-session reuse.
• Update the query-caching and model-catalog-compat OpenSpec contracts.

Test plan

PYTHONPATH=. pytest \
  tests/unit/test_cache_invalidation_poller.py \
  tests/integration/test_cache_invalidation_bus.py \
  tests/integration/test_upstream_route_cache_invalidation.py \
  tests/integration/test_model_registry_replication.py \
  tests/integration/test_multi_replica.py -q -x
# 74 passed

ruff check <touched Python files>
ruff format --check <touched Python files>
ty check --python .venv <touched Python files>

openspec validate recover-cache-invalidation-after-prime-failure --strict
openspec validate --specs --strict
# 48 main specs passed

The full repository-wide pre-commit local-ci gate was intentionally not run
locally; the focused Sensitive routing/cache suite, scoped lint/format/type
checks, strict OpenSpec validation, and independent diff review were completed.
Repository CI remains the authoritative integration gate.

Screenshots / output

Not applicable. This changes backend poller lifecycle behavior and has no
dashboard-visible surface; the real routing-cache and bridge-reuse seams are
covered by integration tests.

Checklist

• Title is in Conventional Commits format (<type>(<scope>)?: <subject>).
• Linked the related issue where applicable; no matching issue was found.
• Added or updated tests covering the change.
• Ran the relevant focused local test/lint/type subset.
• If touching specs: strict OpenSpec validation and verification are clean.
• Simplicity gates reviewed: no settings, setup steps, README sections,
  environment examples, dashboard navigation, or UI defaults are added.
• CHANGELOG is not edited by hand.

[mastertyko]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Hooray!

Reviewed commit: 39283463e8

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".