Serve history pages of a resident session from its cached log

[Tryanks]

For a long thread that was only opened (never appended to in this process), every history snapshot and page re-read and re-parsed the whole JSONL, and turn_aligned_start re-folded from record 0 each call. After #454 a 203k-record thread needs 16 pages instead of ~1,000, but each still paid the full read.

Change

• AppState::event_records (the existing per-session record cache) now holds a SessionLog: the records, the pure fold, and the indices of turn-opening records. turn_aligned_start is a binary search over those indices. Envelope, SESSION_HISTORY_RECORDS and paging semantics are unchanged.
• One parse per residency: opening a session (TimelineLoadTarget::Active) loads the log on the mailbox in the same turn the snapshot needs it; the background timeline load clones the cached fold instead of re-reading. Background (orchestrate child) loads keep parsing off the mailbox and install the log on completion. record_event_at still lazily loads on first append.
• A non-resident session is read cold and not retained, so paging it never grows the map.
• Memory policy: a log exists for live and parked sessions only (bounded by the resident LRU). release_stale_session_logs runs at every residency exit; it queues a store-writer flush barrier and drops the log when the barrier echoes, unless the session became resident again or appended meanwhile (re-armed). A cold read after the drop therefore sees every append the log accepted, so the JSONL lagging the writer can never truncate a reopened thread.
• Snapshot total_turns now comes from the cached fold rather than the resident timeline, which was still Timeline::default() while the async load was pending.
• SessionStore::event_reads() counter behind a new tcode-services/test-support feature (resolver 3; dev-dep only, zero cost in production).

Timings (real 25.2 MB / 202,941-record log, snapshot + 16 pages)

	before	after
debug	11.62 s	0.52 s
release	1.08 s	24 ms
The one-time read moved into select_session (73 ms release, 676 ms debug) and the second parse for the timeline load is gone.

Tests

• history_pages_of_an_opened_session_parse_the_log_once (old code: one read per page → fails).
• session_log_follows_residency_and_flushes_before_release (old code never dropped entries → fails).
• history_paging_bench (#[ignore], TCODE_HISTORY_BENCH_LOG=<path>).
• parked_readopt_refolds_events_appended_while_parked now appends through record_event (the production writer) instead of writing to the JSONL behind the runtime; the contract it protects is unchanged.

cargo fmt --all --check, cargo clippy --workspace --all-targets --locked -- -D warnings, cargo test --workspace --locked, cargo machete green.