[codex] cache state DB runtime for repeated lookups

[charley-oai]

What changed

• cache the core state DB runtime for ad hoc lookup helpers like get_state_db() and open_if_present() so repeated cold metadata/path lookups can reuse an existing StateRuntime
• keep state_db::init() creating a fresh runtime for session-owned work instead of funneling everything through one shared SQLite pool
• allow no-provider open_if_present() callers to initialize and cache an ad hoc runtime so rollout-path fallback and DB read-repair still work
• treat empty-provider cached runtimes as reusable for later ad hoc lookups with a configured provider
• prune dead cached runtimes on insert so one-off sqlite homes do not accumulate stale cache keys for the process lifetime
• add focused tests covering cached lookup reuse, no-provider initialization, distinct-runtime behavior for init(), and stale-cache pruning

Why

Cold thread metadata lookups were repeatedly calling StateRuntime::init(), which opens both state_5.sqlite and logs_1.sqlite. On machines with a very large logs_1.sqlite, that made cold thread/read and thread/resume paths pay repeated logs DB open/migration costs even when they only needed thread metadata or rollout path lookup.

The first version of this change cached too broadly. Each StateRuntime uses a SQLite pool capped at 5 connections, so caching state_db::init() as well would have risked routing session-owned state DB work from multiple live sessions through one shared 5-connection pool. This PR keeps the caching benefit for ad hoc lookup-style callers while preserving separate runtimes for session-owned paths.

The cache also stores Weak<StateRuntime> values keyed by sqlite_home. Without pruning, dead entries can accumulate if many distinct homes are initialized once and never revisited. Cleaning them up on insert keeps the cache bounded in those cases.

Impact

• reduces repeated SQLite runtime initialization on cold lookup paths
• preserves filesystem fallback and read-repair when find_thread_path_by_id_str() reaches the no-provider open_if_present() path
• keeps session-owned state DB work on distinct runtimes instead of concentrating it on one shared 5-connection pool
• avoids unbounded growth of dead cache keys across many distinct sqlite homes
• keeps existing backfill-complete gating for regular state DB reads

Validation

• cargo test -p codex-core rollout::tests::find_thread_path_falls_back_when_db_path_is_stale
• cargo test -p codex-core rollout::tests::find_thread_path_repairs_missing_db_row_after_filesystem_fallback
• cargo test -p codex-core state_db::tests
• just fmt
• just argument-comment-lint
• cargo test -p codex-core still shows unrelated pre-existing failures in other areas; the rollout regressions above are fixed

Notes

• I did not run the full workspace test suite because repo guidance says to ask before doing that for core changes.

[charley-oai]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0d3e0fb47b

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".

[charley-oai]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Bravo.

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".