fix(cli): doctor ledger lock, catalog timeout and host check; sibling watcher starts at the log end

[ajianaz]

What

Follow-ups from the cora 0.19.0 agent-reviewer scan (the doctor and sibling findings, verified against the code) plus the three guarded unwraps cora reported.

Why

• Concurrent tole doctor runs could overwrite each other's ledger entry (unlocked load/record/save) and race on the temp file.
• The catalog tier reused the ACP model picker's 3 s probe timeout, so a slow but healthy gateway showed up as a release-gate blocker.
• base_url_ok accepted URLs without a host (https://?x).
• Every child agent folded the whole shared touch log (all earlier children's history) on start-up.

Changes

• doctor.rs: Ledger::update holds an exclusive advisory flock on doctor-ledger.json.lock (owner-only) across load/record/save, the same pattern as the agents-root lock (no lock on non-unix, single writer). The catalog tier uses CATALOG_TIMEOUT = 20 s. base_url_ok requires a host after any userinfo and before /, ?, #.
• openai.rs (tole-core, additive): fetch_model_ids_with_timeout; fetch_model_ids keeps its 3 s behaviour for the picker.
• sibling_watch.rs: SiblingWatcher::from_end starts at the last line boundary (reads only the tail; a half-written last line is read once complete). tole-cli uses it for the child watcher.
• main.rs, session_host.rs: if let Some(search) = WebSearchTool::from_env() instead of is_some() + unwrap().
• CHANGELOG.

Testing

• New tests: concurrent ledger updates from 8 threads keep all 8 entries (fails 3/3 with the lock removed, passes with it); base URL host matrix; from_end skips history but sees later activity, and waits for a half-written last line.
• fmt, clippy -D warnings, cargo test --workspace, cargo check --no-default-features (core and cli) pass; cora review --staged clean.

[github-actions]

🔍 Cora AI Code Review

✅ No issues found. Code looks good!

---

Review powered by cora-code · BYOK · MIT