73. Dependency-Component Propagation of Variable Reads

• ruleVariables previously ran a fresh reachable-rule DFS per rule, collecting the same transitive reads repeatedly, even when no rule read variables. It now collects direct reads lazily and returns an empty result before constructing a graph when there are none. Otherwise, the existing syntactic call traversal and Tarjan ordering place callee components first. Each component unions its direct peer reads and finalized external callee keys once, sorts the result, and shares that immutable slice between recursive peers. Missing callees contribute nothing. Definitions alone are not reads; recovery arguments and Operators Rule skip/operand/operator expressions/actions keep their existing conservative dependency contract. Memo-key values and matching behavior are unchanged in engines and standalone generated parsers.

• Current-parent compilation of a 512-rule no-variable chain reproduces 10.212 ms / 13.400 MB per operation in a one-second profiled diagnostic. An allocation profile attributes 87.61% of sampled bytes directly to ruleVariables (87.68% cumulative), confirming the earlier compiler audit. This is an allocation attribution, not a CPU fraction; compilation includes all phases. Profile totals also include benchmark setup.

• Five alternating parent/candidate pairs with a 200 ms benchtime compare baseline a1424e5 with component propagation on Apple M3 Max, Go 1.27.1. Grammar parsing and fixture generation are outside timing; full Compile with default options is measured. Chains have the listed generated rules plus main, in either declaration order; a leaf action either reads x or returns a constant. The main rule defines x in both controls.

  Chain rules / reads	Caller-first time before → after	Callee-first time before → after	Caller-first B/op before → after
  64 / none	0.250 → 0.117 ms (−53.3%)	0.246 → 0.108 ms (−56.2%)	407.55 → 214.16 KB
  64 / leaf variable	0.260 → 0.137 ms (−47.5%)	0.255 → 0.142 ms (−44.5%)	419.39 → 259.24 KB
  512 / none	10.114 → 1.003 ms (−90.1%)	10.063 → 1.017 ms (−89.9%)	13.414 → 1.711 MB
  512 / leaf variable	10.516 → 1.219 ms (−88.4%)	10.521 → 1.223 ms (−88.4%)	13.496 → 2.089 MB

  All eight time ranges are disjoint. At 512 rules, caller-first allocation counts fall 17,466 → 10,196 without reads and 17,658 → 11,789 with the leaf read. Byte/alloc counts are cumulative per compilation, not retained heap.

• Shared and cyclic 128-rule graphs with three leaf reads take 1.250 → 0.490 ms (−60.8%) and 1.764 → 0.378 ms (−78.6%). Graphs with a different direct variable at each of 128 rules take 1.941 → 1.209 ms (−37.7%) for a chain and 3.923 → 0.406 ms (−89.7%) for a cycle. All four ranges are disjoint. The latter cycle drops 4.238 → 0.590 MB and 9,507 → 6,607 allocations; sharing the finalized SCC keys avoids allocating one equal slice per peer. This does not promise linear total set storage: different acyclic closures can still contain many variables, and sorting and other compiler analyses remain.

• A first reverse-edge worklist reduced common cases but repeatedly merged growing sets. Separate 200 ms diagnostic samples with distinct direct reads showed chain/cycle times 1.902 → 6.853 ms and 3.796 → 17.911 ms at 128 rules (about 3.6×/4.7× slower). That design was rejected. Final component ordering resolves cycles together and finishes callees before callers, avoiding those repeated updates. These prototype diagnostics are separate from the final five-pair estimates above.

• The unchanged calculator/minilang/outline/JSON grammar AST controls change −5.8%/−17.3%/−7.1%/−18.6% in full compilation time. Calculator/outline ranges overlap; minilang/JSON do not. Bytes decrease approximately 1.0%/9.7%/0.7%/12.7%, and allocation counts fall 11/240/3/116, respectively. No parsing-runtime speedup is claimed. Positive-read analysis still builds transient call maps and SCC metadata; the measured benefit comes from avoiding repeated transitive DFS and evolving-set propagation.

• Exact sorted sets match an independent per-root DFS oracle for empty, missing-reference, reordered 64/512/2,048-rule chains, shared/cyclic graphs, and shuffled distinct-variable growth graphs. Traversal checks include choices, lookaheads/discard, optional/repetition, recovery sync, Operators Rule actions, self calls and bindings with no reads. Existing variable-sensitive memo tests cover engine backends/position units, memo on/off and Document edits; generated-runtime and saved-module tests retain the same keys. Independent source review verifies callee-first order, missing/duplicate definitions and read-only use of shared metadata.

• Reproduce with go test ./internal/engine -run '^$' -bench '^BenchmarkVariableDependencies$' -benchtime=200ms -benchmem. Build both binaries with identical benchmark source, restoring only analysis.go from a1424e5 via a Go overlay for the parent. Alternate five pairs with no concurrent tests/builds/probes. For allocation attribution, run BenchmarkVariableDependencies/no-vars/caller-first-512 separately with -benchtime=1s -memprofile=/tmp/pego-compile.pprof, and inspect go tool pprof -sample_index=alloc_space. Raw focused records stay local. The full-suite checkpoint follows this compiler-analysis optimization.