perf(model_metadata): salvage probe-cache cluster

[kshitijk4poor]

Summary

Salvages the remaining worthwhile pieces of the old probe-cache / local-model metadata swarm onto current main.

This PR preserves and integrates:

• fix(model_metadata): cache detect_local_server_type result for process lifetime #29988: process-lifetime cache for detect_local_server_type()
• perf(model-metadata): resolve localhost to IPv4 in detect_local_server_type #37595: localhost→IPv4 rewrite to avoid Windows dual-stack IPv6 timeout stalls
• fix(model_tools): honor model.context_length to skip OpenRouter probe on banner #50572: honor configured model.context_length on the banner path to skip unnecessary OpenRouter probing
• perf(provider): cache model endpoint probes to avoid 5s delay on every query (fixes #41997) #42081 kernel, reworked: 30s positive-only TTL cache for _query_ollama_api_show()
• fix(model_metadata): eliminate idle Ollama API probes from gateway processes #37905 kernel, reworked: trailing-slash normalization for persistent context-cache keys, with legacy-key read/invalidate compatibility

Why this salvage shape

A straight resurrection of any single old PR would either miss sibling sites or carry stale behavior.

#37595 widened to sibling probe sites

The original localhost→IPv4 rewrite fixed the timeout only inside detect_local_server_type(). The same localhost -> ::1 -> connect timeout -> IPv4 fallback penalty existed at the sibling probe helpers that build URLs directly from base_url, so this salvage extracts _localhost_to_ipv4() and applies it consistently across:

• detect_local_server_type()
• query_ollama_num_ctx()
• query_ollama_supports_vision()
• _query_ollama_api_show()
• _query_local_context_length() (including the LM Studio native API URL)

That fixes the bug-class instead of just one entrypoint.

#42081 reworked to positive-only caching

The remaining hot-path gap on current main was _query_ollama_api_show(): repeated back-to-back resolutions could still issue repeated POSTs. This salvage adds it to the existing in-process TTL probe cache, but intentionally caches only positive results.

We do not memoize failures: if a local server is still starting up, a transient failed probe must not suppress a legitimate retry seconds later.

#37905 reworked for backward compatibility

Persistent context cache keys are normalized through _context_cache_key() so http://host/v1 and http://host/v1/ share a single row. Reads and invalidation also honor legacy pre-normalization rows so old cache files do not force unnecessary re-probes.

Validation

Targeted tests:

• tests/agent/test_probe_cache_followups.py
• tests/agent/test_model_metadata_local_ctx.py
• tests/agent/test_model_metadata.py

Result:

• 163 passed, 0 failed

Mutation check performed on the positive-only guard:

• intentionally changed _query_ollama_api_show() to cache failures unconditionally
• test_failure_never_memoized failed as expected
• restored the guard and re-ran green

Authorship / credit

Preserved contributor commits where still valid:

• @uzaylisak from fix(model_metadata): cache detect_local_server_type result for process lifetime #29988
• @rodboev from perf(model-metadata): resolve localhost to IPv4 in detect_local_server_type #37595
• @zhchl from fix(model_tools): honor model.context_length to skip OpenRouter probe on banner #50572

The #42081 and #37905 pieces are salvaged as current-main reimplementations of the still-useful kernels rather than direct cherry-picks, because the original forms were either behaviorally wrong for Hermes current startup races (#42081 negative caching) or incomplete / stale relative to main (#37905).

Phase-2 gates (post-open review)

Structured 2a/2b/2c gates were run on this PR; findings folded into the follow-up commit:

• 2a tests: targeted suites 202 passed / 0 failed; full tests/agent/ has 17 failures that are identical on clean upstream/main (env-dependent anthropic/bedrock/credential-pool tests, pre-existing); ruff clean; mypy delta vs base: 0 new errors.
• 2b live smoke: 6/6 PASS — process-lifetime detect cache, one-POST TTL behavior, failure-not-cached, disk key normalization, IPv4 rewrite passthrough, config context_length honored end-to-end (54321 from a temp config).
• 2c review findings, all fixed:
  • detect_local_server_type memo is now a bounded 1h TTL instead of process-lifetime, so a backend swap on the same port re-detects.
  • legacy disk-row compatibility now works in both migration directions (lookup and invalidation).
  • _localhost_to_ipv4 is scheme-anchored host-only (embedded http://localhost in a query string no longer corrupted).
  • disk invalidation also drops the in-memory TTL rows, so a stale value can't be re-persisted within the TTL window.
  • AUTHOR_MAP entry added for the fix(model_tools): honor model.context_length to skip OpenRouter probe on banner #50572 cherry-pick author email (CI attribution gate).