VCR-RA-010: const-remat eligibility DERIVED from the stream, the hint only gates (#242)

[avrabe]

Wave-2 lane L2 of the v0.55 "scry seam" hub. Flag-off; bytes unchanged.

scry's interval domain identifies locals with a singleton interval [c, c] — provably constant values the allocator can rematerialize at the use instead of spilling and reloading. That signal was a dev-dependency test (scry_const_remat_signal.rs); this promotes it to an allocator input.

The design question

A wrong --proven-safe verdict (#901) elides a bounds guard — a memory-safety hole. A wrong const-remat verdict would re-emit the wrong value — a miscompile. So this seam follows the --wcet-hints shape, not the trusted-fact shape:

Eligibility is DERIVED from the emitted ARM stream. The hint only GATES consumption and can never create an eligible site.

plan_const_remat walks the final stream and admits a reload only where it can itself see the constant definition reaching the spill with no intervening write. MOVT over an unknown low half stays unknown; a non-constant store retires the slot; a branch/call/label closes the window.

Conservatism is reported, not hidden

Any op the walk cannot classify clears all state — a large amount of missed opportunity, deliberate for a first increment, because an op wrongly treated as harmless could hide a write to the slot and turn a stale constant into a miscompile. reloads_seen and windows_closed accompany every site count so a small number reads as narrow window, never as small opportunity. Widening the safe set op-by-op is the named follow-up.

Mutation earned its keep

Mutating the walk to treat an unknown register as the constant 0 exposed that the red-first test lying_hint_yields_no_candidates passed for the wrong reason: it offered hint value 999 while a broken walk invents 0, so the gate's value comparison did the rejecting, not the stream check. Fixed to run the gate OPEN so only the stream can reject — it now goes red under that mutation, as do two other tests.

7/7 tests, clippy clean, cargo fmt clean.

Refs #242

[codecov]

Codecov Report

❌ Patch coverage is 94.01408% with 17 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
crates/synth-synthesis/src/const_remat.rs	94.01%	17 Missing ⚠️

📢 Thoughts on this report? Let us know!