mutation survey: DEAD means unreached by a soft-float, flag-off corpus — all four v0.65 DEAD sites are reachable (RQ-66-DELETE)

[avrabe]

Finding (RQ-66-DELETE, epic #242)

RQ-65-MUTANTS classified four sampled mutants DEAD — byte-identical on the corpus AND the site never evaluated — and v0.66 scoped them as the first principled deletion list for the subtraction ratchet. Before deleting, RQ-66-DELETE tried to REACH each site instead. All four are reachable. Under the survey's own three configurations every probe stays silent (the v0.65 verdict reproduces exactly); under configurations the corpus never compiles, every site fires:

site (crates/synth-backend/src/arm_backend.rs)	survey (3 x cortex-m4)	reached by
:808 if vfp.is_ok() {	unreached	vfp_spill_881.wat, vfp_local_pressure_1069.wat on --target cortex-m7dp (reloc + self) and cortex-m4f
:836 || msg.contains("spilling the VFP register file")	unreached	vfp_local_pressure_1069.wat (live24, the grown-pool composition) on the same three hard-float configs
:862 if grown.is_ok() {	unreached	vfp_local_pressure_1069.wat on the same three
:1316 if literals > 0 {	unreached	171/203 modules with SYNTH_GRAPH_ALLOC=1 --relocatable, 108/203 self-contained

Measured with the survey's OWN probe ({ eprintln!(MARK); token } on the original token, Edit(site, "probe")), the survey's own enumerate_sites() and corpus, and a rebuilt synth per site; tree restored and asserted clean after each.

Root cause — a property of the corpus configuration, not of the code

• CORPUS_CFGS compiles every module as --target cortex-m4 (no FPU). The GI-FPU-002 VFP retry ladder (GI-FPU-002 + RA tail is now the ONLY gate between the falcon cascade and the M7 — 5 entry-point symbols; phase-2 D-register pressure is new in v0.52 (inline-f64 #869 lowering) #881 / GI-FPU-002: three named cascade entry points are the entire remaining gap to a complete falcon M7 image (attitude#tick, ekf#estimate, position#tick) — everything else in the chain now works #1069) can only be entered on a hard-float target, so its three rungs are unreachable BY CONSTRUCTION under the survey — while the two fixtures written to exercise that ladder sit IN the corpus and are compiled by their own CI oracles on cortex-m7dp.
• The graph-colouring arbiter (literals > 0 sizing) lives behind the flag-off SYNTH_GRAPH_ALLOC spike; the corpus sets no environment, so nothing under graph_alloc::enabled() can ever be reached by the survey, although the vcr_dec_001_graph_alloc_differential job runs it on every PR.

So DEAD = "unreached by a soft-float, flag-off corpus". That is exactly what the report says it means; it is not a deletion list.

A second structural fact

selector_lines_code (the ratchet) counts crates/synth-synthesis/src/instruction_selector.rs + instruction_selector/**. All four DEAD sites are in arm_backend.rs, and four of the survey's five regions (R1 routing, R2 optimizer_bridge.rs, R4 shared tail, R5 startup) are outside that population — only R3 (select_with_stack.rs) is inside it. A DEAD site from R1/R2/R4/R5 can never move the ratchet even if deleted.

What RQ-66-DELETE ships

• scripts/mutation_survey.py reach [--write]: re-probes DEAD sites under REACH_CFGS (hard-float targets, SYNTH_GRAPH_ALLOC=1) and records the reaching modules per configuration on the ledger record (reach_wide).
• pin-subset carries want_reach_wide witnesses; ci requires every pinned witness to keep reaching (a deletion, or a probe that went blind, is red) and prints MUTANTS-REACH-WIDE entries=N reached=N unreached=0, grepped in the mutation-survey-discrimination job.
• The DEAD classification is left as recorded (it is true for the stated frame); the report gains a section stating all four are reachable and are not deletion candidates.

Still open after this

• The corpus configurations themselves. Adding a hard-float and a flag-on configuration to CORPUS_CFGS would change every baseline hash and changed set in the ledger — that is a re-survey (RQ-65-MUTANTS item 3 in the report), not a patch, and should be done when the survey is next re-run with the parity oracle in the suite.
• Any future DEAD verdict should be read as "unreached under CORPUS_CFGS + REACH_CFGS" and still needs a reachability argument before deletion.

Refs #242

[avrabe]

Addressed in PR #1239 (RQ-66-DELETE): the reach gate (REACH_CFGS, reach subcommand, want_reach_wide witnesses replayed by ci, classify_identical so a wide-only reach is UNRESOLVED and never DEAD), the four ledger records annotated with published_as (v0.65.0's "4 DEAD deletion candidates" is false 4 of 4), and the report's follow-up section. Left open: how the v0.65 correction is published (coordinator's decision), folding REACH_CFGS into the byte-triage baseline (a re-survey), and deriving the configuration set from the levers the code generator reads instead of hand-listing it.