71. Dependency Propagation for Sample Constructor Analysis

• sample.New previously swept every rule until minimal derivation heights, byte lengths and coverage reach sets stabilized. Caller-first alias chains improved one link per sweep. Analysis now builds reverse dependencies, seeds a bounded deduplicated worklist in explicit DFS postorder, and reevaluates only callers of changed estimates or sets. Acyclic callees run first; cycles still converge to the previous fixed point. Repeated calls add one reverse edge, stored in a contiguous array with prefix offsets. Target numbering, generation decisions and negative-lookahead/recovery exclusions are unchanged. This affects the shared sampling constructor, which requires a grammar AST, rather than matching in any backend or generated parser.

• Five alternating 200 ms pairs compare baseline 9159f94 with this dependency propagation on Apple M3 Max, Go 1.27.1. Grammars are parsed before timing; only analyze(g, "main") runs in the timed loop. Each chain ends in "é" and has the listed derivation depth (one more rule than the depth):

  Depth	Caller-first time, before → after	Leaf-first time, before → after	B/op, before → after
  128	0.502 → 0.0391 ms (−92.2%)	0.0361 → 0.0378 ms (+4.7%)	71.94 → 81.32 KB
  512	9.687 → 0.1680 ms (−98.3%)	0.1554 → 0.1603 ms (+3.2%)	316.51 → 355.55 KB
  1,024	44.152 → 0.4001 ms (−99.1%)	0.3637 → 0.3783 ms (+4.0%)	802.42 → 878.71 KB
  2,048	234.498 → 0.9909 ms (−99.6%)	0.9525 → 0.9934 ms (+4.3%)	1.795 → 1.949 MB

  Caller-first ranges do not overlap; all four leaf-first ranges overlap. All chains add nine allocations, irrespective of depth, and B/op grows 8.6–13.0% for transient dependency/queue storage. An initial per-rule caller slice design added 136–2,056 allocations; contiguous edges avoid that cost. At depth 2,048 it used 1.988 MB versus the final 1.949 MB. Preserve the simpler fixed-point semantics while trading this bounded startup storage for avoiding repeated whole-grammar work. Dense per-rule reach bitsets remain unchanged; their memory still scales with rules times targets. Sparse/shared sets need separate measurements and ownership design before changing representation.

• The same pairs include five controls: a mixed graph fixture with shared calls/finite and impossible cycles, a one-rule choice/repetition, and the unchanged calculator/minilang/outline example ASTs. Median analysis time changes −2.1%, +8.7%, −3.9%, −20.0%, −14.1%, respectively. The first three sample ranges overlap; minilang and outline ranges do not. Controls add 4–7 allocations and 2.5–9.2% B/op. The one-rule median rises roughly 0.655 → 0.712 µs. These are constructor-analysis results, excluding source compilation, generation attempts and parsing; they do not claim a runtime parsing speedup or a limit on constructor work from WithBudget.

• Reordered 128/2,048-depth chains and 30 shuffled shared/recursive graphs agree with an independent whole-sweep scheduling oracle for every rule's height, length and reach set. Tests preserve impossible/missing-reference estimates, generation-specific call exclusions, Unicode byte lengths and exact chain coverage. Existing deterministic/valid/invalid generation, dead-alternative, Operators Rule-level and engine-corpus checks pass. A fixed-seed public generator probe produces byte-identical inputs, coverage report and statistics to the parent on the same Go release. Independent source review checks CSR count/fill, cursor-buffer reuse, dependency completeness and bounded ring invariants.

• Reproduce with go test ./sample -run '^$' -bench '^BenchmarkSampleAnalysis$' -benchtime=200ms -benchmem. Build parent/candidate binaries with the same benchmark source; restore only sample/analysis.go from 9159f94 through a Go overlay for the parent. Run five alternating pairs with no tests/builds/probes concurrently. Raw results stay local. The full-suite checkpoint follows the current core compiler analysis optimization group; batch-runtime benchmark workloads are unchanged.