feat(chunk): add pre-encoded SourceMapChunk and stitch_chunks (esbuild model)

[Boshen]

Summary

First step toward an esbuild-model source map pipeline: build the mappings VLQ once and join maps by stitching pre-encoded bytes, instead of materializing a combined Vec<Token> and re-serializing.

SourceMapChunk mirrors esbuild's sourcemap.Chunk — a map's mappings already encoded to VLQ plus the small delta-carry state needed to splice it. stitch_chunks re-encodes only each chunk's first segment (and first tail name) against the running state, then copies the rest of the bytes verbatim → O(modules) work, not O(tokens).

API

• SourceMapState — delta-carry state of a mappings stream
• SourceMap::to_chunk() — encode mappings once, capture seam metadata
• stitch_chunks(&[(&SourceMapChunk, line_offset)]) — O(modules) join of pre-encoded chunks
• encode_vlq_diff made pub(crate) to reuse the exact VLQ primitive

Correctness

stitch_chunks output is asserted byte-identical to the trusted path ConcatSourceMapBuilder::from_sourcemaps(...).into_sourcemap().to_json().mappings:

• 4 curated cases (single chunk; no names; name in first-segment vs tail; 3-way mixed)
• a 2000-iteration deterministic fuzz (seeded LCG, no rand dependency) over random module counts, token streams, names, and line offsets

Notes

• Invariant: each chunk's first token must carry a source (a code generator always emits a start-of-file mapping; esbuild assumes the same). The fuzz respects it.
• ConcatSourceMapBuilder already appends names/sources wholesale (no cross-module dedup) — same as esbuild — so the stitch is byte-identical with no names tradeoff.
• Infrastructure only / behavior-preserving: nothing produces SourceMapChunk yet. Next: a Mappings::{Tokens,Encoded} representation so to_json_string short-circuits and SourceJoiner stitches, then route no-chain modules through to_chunk.

Draft — opening for early review of the stitch math and API shape.

[codspeed-hq]

Merging this PR will degrade performance by 2.03%

❌ 2 regressed benchmarks
✅ 14 untouched benchmarks
⏩ 7 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
❌	lookup_table[real_small]	1.3 µs	1.4 µs	-2.13%
❌	lookup_table[real_medium]	1.5 µs	1.5 µs	-1.93%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing perf/sourcemap-chunk-stitch (19f9868) with main (296f357)}

Footnotes

1. 7 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩