fix(coding-agent): avoid transcript scans for exact session IDs

[metaist]

@davidbrai
Fixes #9440

• Uses an exact session-ID lookup to read session headers instead of loading the whole transcript. I did compare keeping the async contract vs doing it sync and the latter was better for subsequent cache hits, so I dropped the async bits.
• Quick microbenchmark showed ~19.5x reduction in lookup time (68.1s -> 3.5s). Still linear in the number of transcripts, but tolerable for now.

Ran npm run check and ./test.sh which passed.

[holny]

Checked the diff against the existing discovery code: the filterCwd condition in findById matches list() (session-manager.ts:1672) and continueRecent() (:1591) verbatim, so exact-id lookups in a custom --session-dir stay consistent with how those filter by cwd. readSessionHeaderForDiscovery is the same bounded header scan findMostRecentSession() uses, and files exceeding the scan limit are skipped exactly like in list(), so corrupt/oversized sessions behave the same as before. The vi.spyOn(SessionManager, "list") in the regression test is a neat way to pin the no-full-listing contract.