fix: Shared-shell replay shifts 1 MiB on every output chunk

[sam-saffron-jarvis]

What changed

• Keep a logical start index for the shared-shell replay buffer instead of shifting the retained 1 MiB window after every small PTY output chunk.
• Reclaim dropped storage only after a full replay window has accumulated, while preserving baseOffset and nextOffset semantics.
• Account for the logical prefix in chunked snapshots and collaboration protocol redaction/replacement.
• Add a regression test that crosses multiple batched compactions with 32 KiB chunks and verifies exact retained bytes, offsets, reset behavior, redaction, and chunked replay.
• Add a 256 MiB sustained-small-chunk benchmark.

Why this is high-value

The Unix PTY producer reads up to 32 KiB at a time. Once the replay buffer reached 1 MiB, each ordinary output callback previously copied the entire retained window while holding the shell mutex. That meant about 32 MiB copied for each 1 MiB of continued output, delaying snapshots, browser input, lifecycle operations, and collaboration state during verbose builds, tests, searches, and log streams.

The logical prefix makes eviction constant-time per chunk and performs one roughly 1 MiB compaction per additional 1 MiB of output instead of one per 32 KiB chunk: 32x fewer replay-window shifts in the steady producer case. The end-to-end append benchmark improved from a median 520.8 ms (515 MB/s) to 425.3 ms (631 MB/s) for 256 MiB on an i9-14900K, while replay storage remains bounded to roughly two windows between compactions.

Validation

• go test ./cmd -run '^TestServeShellReplay(ResetAndExit|CompactsDroppedPrefixesInBatches)$' -count=1
• go test -race ./cmd -run '^TestServeShellReplay(ResetAndExit|CompactsDroppedPrefixesInBatches)$' -count=1
• go test ./cmd -run '^$' -bench '^BenchmarkServeShellReplaySmallChunks$' -benchtime=1x -benchmem -count=3
  • Before median: 520768260 ns/op, 515.46 MB/s
  • After median: 425333161 ns/op, 631.12 MB/s
• go build ./...
• go vet ./...
• go test ./... ran successfully for the changed cmd package and all other root-module packages except an unrelated existing generated-frontend budget failure in internal/serveui: app.js gzip 138137 > 135000 and app.css gzip 33383 > 33000. This branch changes no frontend source or bundle-budget code.

[SamSaffron]

Closing at the maintainer’s request in favor of a consolidated implementation on the current working tree (not merged or pushed yet). Retained as an 8/10 priority: batch shared-shell replay compaction rather than copying 1 MiB for every small PTY chunk. Includes offset/redaction/chunked replay coverage and an additional large-write-after-compaction regression. Full Go suite, relevant race tests, build, vet, and diff checks pass.