v0.3.0 — chunked materialize for huge forks

The fix

Forking a huge session (60万+ events) crashed or OOM'd because encodeMaterialization → eventLines = map(JSON.stringify).join() builds a single ~501MB string; 74万 events throws RangeError: Invalid string length (V8 ~512MB cap).

v0.3.0 patches JsonlSessionPersistence.encodeMaterialization (rc.6 signature-guarded) to emit one zstd frame per 50k events. Multi-frame output is natively decoded by dsh's scanZstdFrames, byte-compatible.

	heap peak	74万 events
before	1572MB	RangeError
after	1155MB	safe

Why 'single session fine, fork breaks'

• normal sessions persist incrementally (appendLines, small batches)
• fork children persist via materialize (full seed) — the only path that serializes the whole log at once

Install

dsh plugin --profile web add github:orangeofcarl0-sys/dsh-projection-warmup

v0.2.0 — chunked warmup + fork backfill

What

• chunked live-session warmup (default on): folds projection cells in chunks with setImmediate yields, writing them into the registry WeakMap so snapshot()/drive() hit warm cells — a 74万-event session drops from a 20-min cold fold to ~200ms.
• fork child cache backfill (default on, zero-cost): fork children get projection cache rows written right after warmup, so abandoned forks never hit the full readFrom(0) coldSnapshot path.
• disk cold-session backfill (default OFF): streaming parse+fold to write missing cache rows. Off by default because readRaw's sync zstd decode is not chunkable at plugin layer.

Install

dsh plugin --profile web add github:orangeofcarl0-sys/dsh-projection-warmup

Verified on dsh 0.1.0-rc.6 / Windows (tests: 19 asserts ALL PASS).