feat(server): improve operational diagnostics and writer tuning

[XiNian-dada]

Summary

• clarify that audit degradation is diagnostic while history and registry remain readiness hard dependencies
• expose production token-cache hit, miss, and capacity-eviction counters through Prometheus
• make history and audit writer batch limits and flush intervals configurable end to end

Details

Readiness semantics (#313)

• keep HTTP 200 and ready: true when only audit/write diagnostics are degraded
• keep HTTP 503 for history or registry hard-dependency failures
• return explicit status: degraded and problems details
• document probe behavior in docs/readiness.md and add both HTTP/JSON regression combinations

Token cache diagnostics (#306, Phase 1 only)

• export nodelite_token_cache_hits_total
• export nodelite_token_cache_misses_total
• export nodelite_token_cache_evictions_total
• count both initial and post-semaphore lookups as hits, count only actual Argon2 work as misses, and count only capacity-driven LRU removals as evictions
• cover warm hits, cold misses, capacity pressure, and concurrent verification semantics

This PR intentionally implements instrumentation only. It does not close the remaining #306 benchmark, root-cause analysis, or hit-rate follow-up work.

This supersedes PR #307: it carries the diagnostic intent into production Prometheus metrics, adds eviction accounting and concurrency coverage, and is fully formatted. PR #307 only exposed test/load-test hit and miss counters and its Lint and Test job stopped at cargo fmt --check.

Writer scheduling configuration (#235)

• add history and audit batch/flush settings across raw config, defaults, public config, validation, example TOML, installer install/upgrade paths, and fixtures
• preserve defaults of 128 records and 100ms
• allow batch sizes from 1 through 4096 and reject values outside that range
• reject flush intervals below 10ms
• defensively clamp both writer batch thresholds and vector reservations even when callers bypass parsed configuration
• wire configured values into both writer contexts
• verify batch-triggered and timer-triggered persistence directly through SQLite without query/shutdown forced flushes

Verification

• cargo test --workspace
• cargo +stable fmt --all --check
• cargo clippy --all-targets -- -D warnings
• sh scripts/test-install-server-config.sh

Closes #313
Refs #306
Closes #235

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 566d8e45a4

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".