Fold the rule's directory into command identity (fixes #167)

[typeless]

Fixes #167.

The bug

Command text is Tupfile-relative, so the same rule in sibling directories renders the same string. compute_command_identity hashed that text plus sticky Variable values and nothing else, so two distinct rules shared one identity. find_by_identity is a lower_bound, so one twin silently won every lookup.

Identity is the cross-build join key for expand_implicit_deps, preserve_old_implicit_edges, merge_out_of_scope_commands, remove_stale_outputs, detect_new_commands and reconcile_input_set — the collision misrouted all of them.

Lost header dependencies, in a plain full build. Two directories compiling a same-named source with the same command line is an ordinary layout:

b's command text	b/main.o after editing b/dep.h
identical to a's	unchanged — rebuild silently lost
distinct (-O0 added)	recompiled, correct

No scoping, no globs, no -B. Byte-identical object after the header changed, with no diagnostic.

Index truncation. A scoped build dropped every out-of-scope command, because merge_out_of_scope_commands saw the twin's identity already present and concluded the command was carried forward. Afterwards a full build detected nothing at all — not out-of-scope edits, not entirely new directories. That was #167 as originally filed; it turned out to be a symptom, and the -B immunity I first reported was a coincidence of the repro (variant builds render distinct %o paths, so the text differs).

The fix

Fold the rule's source directory into the hash. The join key must denote which rule this is, and a Tupfile-relative rendering under-determines that — the hash was missing a term of the rule's definition. Identity still deliberately excludes the input set; #166 handled that side separately.

INDEX_VERSION 14 → 15. Without the bump every stored identity stops matching at once, so remove_stale_outputs would delete every output and rebuild it; the bump discards v14 indexes cleanly instead. Verified by patching a live index's version byte back to 14 — one rebuild, then settles.

The version entry also records that v14 predates the OrderOnly edges persisted for #166. That content change shipped without a bump, and this gate is what closes it — so a future revert of this PR must not silently drop it.

From the systemic review of this change

• reconcile_input_set no longer admits LinkType::Group. Group edges run file → group, and the group → command hop is a separate OrderOnly edge, so find_command_by_id rejected every one: the arm was dead code that read as coverage. Removed and filed as Group-member removal is not routed: reconcile_input_set walks one hop, group membership needs two #169 rather than replaced with a two-hop walk I could not demonstrate fixing anything.
• merge_out_of_scope_commands' header claimed it copies "Normal/Sticky edges". It has carried everything except Implicit since Globs re-expand only on Tupfile content change, so newly added files are invisible #166, OrderOnly included — and that one is load-bearing for removal routing after a scoped build.

Filed separately (all pre-existing, none fixed here)

• Command identity injectivity is assumed by every consumer but never enforced #170 — identity injectivity is assumed by every consumer and never enforced. build_identity_map already sorts, so adjacent-equal detection is ~free and would have made this bug an error instead of a stale object.
• Glob results are sorted by StringId handle, so %f order (and command identity) follows readdir order #171 — glob results sort by StringId handle, not path, so %f order follows readdir order. Measured: putup's %f matches ls -U exactly, never ls. Command lines are not reproducible across machines, which undercuts the premise of the identity hash.
• remove_stale_outputs runs after all detection, so its own deletions are never routed to consumers #172 — remove_stale_outputs runs after all detection, so its own deletions are never routed to consumers. One build exits 0 with a stale output; the next heals it.

Tests

Both observed RED before the fix:

• unit (test_graph.cpp) — same text, different source_dir, must differ. RED showed byte-identical hashes.
• E2E (identity_collision) — two directories with byte-identical Tupfiles and sources; editing only b/config.h must move b/program to "Version 2" while a stays at 1. RED showed b stuck at 1.

Verification

615 cases / 142,305 assertions / 32 shards; format, tidy, iwyu clean; no-op stability holds; no bootstrap drift. The original #167 repro now keeps the index at 11 files/2 commands where it fell to 7/1, sees the out-of-scope edit and the new directory, and settles.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HDzr9JV1ofTZKCL1H2qDPg

[github-actions]

PR metrics

Performance (gcc example, Linux)

Workload	Instructions	CPU time	Page faults	D1 miss	LL miss	Wall	Peak RSS
parse	1073 M	0.48 s	14.1 k	0.8% (+0.2pp)	0.1% (+0.1pp)	0.506 s	30.6 MB (-0.2MB)
dry-run	1785 M (+0.2%)	0.58 s	16.6 k	0.9% (+0.2pp)	0.1% (+0.1pp)	0.61 s	40.1 MB

Deterministic signals: instructions (cachegrind-simulated instruction reads — exact across runs, no PMU needed), page faults, peak RSS, and the cachegrind D1/LL miss rates. CPU time is user+sys from time(1).

Internal statistics (gcc example, up-to-date dry run)

Metric	Value
Tupfiles parsed	24
Commands	3545
Commands scheduled	0
Files checked	5818
Files changed	0
Files in index	6087
Graph edges	367913
Index size (bytes)	7264970
Implicit deps	330624
Hash computations	0
Hashes skipped (stat cache)	5818
Stat calls	5869
Parse time (ms)	489.3
Total time (ms)	599.6
Runner CPU	AMD EPYC 9V74 80-Core Processor

Counters from putup -n --stat on the fully-built gcc example (up-to-date dry run): deterministic work measures — a jump in commands scheduled, hash computations, or stat calls is a real behavior change, not noise. Timings are the minimum over repeated runs, compared only against a baseline from the same CPU model; the counters are the regression signal.
Timing deltas suppressed: baseline ran on different hardware (INTEL(R) XEON(R) PLATINUM 8573C).

Binary size (Linux)

Binary	.text	.data	.bss	File
putup	490.6 KB (+0.1%)	1.7 KB	98.7 KB	577.5 KB

Test coverage (lines)

Overall	Median file	Min file	Max file
86.5%	94.9%	0.0% src/platform/path-posix.cpp	100.0% include/pup/core/arena.hpp

103 files · 14983/17313 lines covered

Deltas vs main@787968b7e.

Updated for 9b448ed