perf: skip null-context bind in eachMapping; direct names array read

[7rulnik]

Summary

Three structural changes to the per-mapping eachMapping loop in SourceMapConsumer:

1. Skip aCallback.bind(context) when context === null. Function.prototype.bind produces a bound function V8 can't inline, so every per-mapping call dispatches through the internal [[Call]] path. With no context (the common case — both bench callsites and most consumers omit it), bind is pure overhead and the unbound aCallback inlines normally.

2. Read this._names._array directly instead of this._names.at(idx) per named mapping. Saves the bounds-check + function-call pair — matches the slab-direct pattern already used in _serializeMappings / fromSourceMap.

3. Hoist mapping.source / mapping.name into locals before building the result object. V8 should CSE this, but explicit locals are slightly more inline-friendly.

Bench

SOLO=1 PHASES=eachmapping-generated,eachmapping-original scripts/bench-diff.sh main trace (two-process, ops/sec via benchmark.js):

fixture	order	cand ops/sec	base ops/sec	Δ
amp.js.map	generated	11,648 ±0.52%	3,505 ±8.05%	+232.3%
amp.js.map	original	11,783 ±0.70%	3,489 ±8.07%	+237.7%
babel.min.js.map	generated	1,191 ±0.80%	379 ±1.02%	+214.2%
babel.min.js.map	original	1,143 ±0.52%	383 ±1.07%	+198.4%
issue-41.js.map	generated	1,359 ±0.59%	428 ±1.03%	+217.5%
issue-41.js.map	original	1,366 ±0.70%	429 ±1.14%	+218.4%
preact.js.map	generated	301,487 ±0.30%	68,015 ±0.19%	+343.3%
preact.js.map	original	295,655 ±0.49%	66,782 ±0.29%	+342.7%
react.js.map	generated	106,632 ±0.40%	25,623 ±0.31%	+316.2%
react.js.map	original	89,972 ±0.28%	26,096 ±0.48%	+244.8%
vscode.map	generated	153 ±1.45%	56 ±?	+173.4%
vscode.map	original	150 ±0.76%	54 ±?	+177.2%

Mean +243% across 12 rows.

Caveat on the headline number

The bench passes NOOP as the callback. With NOOP, the per-call cost is dominated by whatever happens outside the callback body, so the bound-vs-direct dispatch gap is fully exposed. With a real callback that does work, the proportional win shrinks (the callback body's cost is unchanged); the absolute time saved per call is the same. Baseline error bars on amp are ±8% vs candidate's ±0.5%, consistent with the bound-dispatch path being in a deopt/reopt cycle that the unbound path avoids.

Test plan

• All 205 tests pass
• Coverage: 98.14 / 97.37 / 96.75 — above thresholds (98 / 97 / 96)