…reaming ReplaceSource

Deep-profiled a webpack-like chunk pipeline (ReplaceSource-wrapped
OriginalSources + SourceMapSources assembled by ConcatSource under
CachedSource) with the V8 CPU profiler and sampling heap profiler.
Three hot spots accounted for ~55% of all allocations; this addresses
each without changing any output:

- createMappingsSerializer built every mapping out of ~5 intermediate
  string concatenations and callers appended the result with
  `mappings += str` (31% of workload allocations). The full serializer
  now writes char codes into a growing Uint8Array via
  createMappingsWriter and materialises the mappings string once in
  finish(). The lines-only variant keeps the string fast paths
  (constant ";AACA" segments + cons append measured faster there) behind
  the same writer interface. The public string-returning
  createMappingsSerializer API is unchanged.

- ReplaceSource.checkOriginalContent lazily split each source content
  into an array of line substrings (splitIntoLines, 11% of
  allocations). It now builds a Uint32Array of line-start offsets
  (exact-size, two-pass scan) and compares with String.startsWith —
  no per-line strings, no comparison slices.

- ReplaceSource.streamChunks now honors finalSource like
  OriginalSource/RawSource: it emits position-only chunks
  (chunk === undefined) and returns the whole replaced source once,
  assembled in O(replacements) string operations by source(). This
  removes the per-token `code += chunk` cons chain in every enclosing
  consumer (ConcatSource, getSourceAndMap, streamAndGetSourceAndMap);
  all in-tree consumers already handle undefined chunks under
  finalSource.

Measured on the profiling workload (interleaved A/B vs main, medians):
total allocations 16.7 -> 6.6 MB/iter (-60%), wall time -14%.
Isolated paths: ConcatSource.sourceAndMap -16..-20% time,
ReplaceSource.sourceAndMap -8%, ReplaceSource.map(columns:true) -8%,
OriginalSource.map allocations 37 -> 9 KB/op at equal time,
columns:false paths at parity. All 89,888 tests pass; sources and maps
are byte-identical to main across the fuzz suite.

Claude-Session: https://claude.ai/code/session_01KDFBEvi3ctDSyHQ7f9izX5

…sions

Addresses the three CI failures on the previous commit:

- lint: reformat the line-start scan loops for prettier 3.9.4 (the
  version pinned by package-lock; a stale local install accepted the
  old shape).

- codecov/project: the legacy lines-only serializer and several writer
  branches lost integration coverage after the writer refactor. Add
  unit tests asserting writer/serializer output equivalence over
  branch-covering and buffer-growth event streams, plus exact-encoding
  cases. createMappingsSerializer.js is back to 100% line coverage and
  overall lib coverage now exceeds the main baseline.

- CodSpeed simulation regressions (benchmarks run under --no-opt, where
  call indirection and extra passes cost far more than JIT-mode wall
  time showed):
  - createLinesOnlyMappingsWriter now inlines the encoding instead of
    delegating to the string serializer per line
    (original-source map({columns:false}) back to parity).
  - checkOriginalContent builds line starts in a single pass with a
    plain array (replace-source streamChunks back below base).
  - ReplaceSource.streamChunks skips assembling the final source when
    the caller passed source: false (getMap does); getSourceAndMap and
    streamAndGetSourceAndMap explicitly override the flag because they
    consume/cache the text — the streamChunks fallback forwards getMap
    options into sourceAndMap, which the fuzz suite caught.

Interpreter-mode A/B vs main after these fixes (medians): replace-source
map() -12%, sourceAndMap() -32%, streamChunks(1000 replacements) -3%,
original-source map({columns:false}) parity, pipeline streamChunks
(columns off) -6%, full-serializer benchmark at parity (its flagged -27%
compares byte-identical code — environment skew, matching the report's
"different runtime environments" warning). Local ground truth for the
flagged webpack-20961 memory rows: allocations 160 MB -> 7.3 MB and
retained heap 42.4 MB -> 3.3 MB in favor of this branch (the writer
produces one flat mappings string instead of a huge retained cons tree).

All 89,895 tests pass; lint and types are green.

Claude-Session: https://claude.ai/code/session_01KDFBEvi3ctDSyHQ7f9izX5