Keep records appended during a timeline load

[Tryanks]

Behaviour

Opening or re-adopting a session inserts the resident with an empty timeline and folds its JSONL on a background task. Events recorded while that fold is in flight were applied to the resident's timeline and then overwritten when the fold landed. The retry-on-watermark loop re-read the file up to four times, but the store writer is asynchronous, so a slow runner could exhaust the attempts and "apply the last fold" without those events. This is the host-side half of the session_replica_matches_live_timeline_for_synthetic_turn failures: in two of the three CI failures the host's own live timeline was missing turn 2 (left had one turn) while the store had recorded it.

The loader now folds from the session's in-memory record cache when one exists (that cache is the complete log once a session has appended in this process; it is only created by record_event_at), otherwise from disk, and on completion applies any records appended past the folded count. The per-process store_append_generation watermark, the attempt counter and the retry are removed; the per-session load generation still discards superseded loads.

Test contract

timeline_load_keeps_records_appended_during_the_fold (renamed from timeline_load_retries_when_append_watermark_moves) records an event immediately after select_session and asserts the loaded timeline contains both the stored and the raced record. It fails at tests.rs:5622 with the catch-up loop removed and passes with it.

Checks

• cargo fmt --all --check
• cargo clippy --workspace --all-targets --locked -- -D warnings
• cargo test -p tcode-runtime --locked (167 + 3 passed)

This does not by itself explain the store-side shape of the flake (replica holding only turn 2); a tracing draft (#433) is running on CI to capture that sequence, and any store change will be a separate PR.