Drew T
2.1 KiB
§401
A probe that does not model the GATE's carve is optimistic, and for jtbl functions it is wrong 92% of the time.
recover_integration --probe-only reports a per-function real cc1 verdict by compiling the draft in
its real TU and comparing that function's bytes. For an ordinary function this is an excellent
predictor. For a jump-table-bearing function it is systematically wrong, because the byte-gate does
something the probe does not: harvest_verify carves the function into its own object so the jump
table can be emitted ([jtbl] carved func_X in its log). The carve changes the translation-unit
boundary, and therefore the codegen — so a draft can match in the uncarved TU and DIFF once carved.
Measured (P31 S70), on 52 leftover standalone-MATCH drafts:
| probe verdict | n | jtbl-bearing |
|---|---|---|
| MATCH (gate says DIFF) | 26 | 24 (92%) |
| CC1-FAIL | 20 | 5 (25%) |
| DIFF | 6 | 0 |
Worked case: ov_SC01_004:func_8017EB30 probes MATCH 279 ins; the gate says DIFF in a worktree and
in-tree and via harvest_verify directly — three independent runs, so it is not a worktree artifact
and not the canon -> cast_call_sites -> sig_unify pipeline either.
Rule. The whole-binary byte-gate is the arbiter (G3). Before believing a probe MATCH, ask whether
the target's .s contains jtbl_[0-9A-Fa-f]{8} (the same predicate harvest_verify carves on, and
parallel_gate._JTBL_RE). If it does, the probe has not modelled the work the gate must do — treat the
verdict as UNPROVEN and do not use it to size a batch. This is the R34 shape: two oracles that disagree
because one is structurally blind, fixed by knowing WHICH is blind, not by trusting the friendlier one.
Corollary — a declaration fix that only turns CC1-FAIL into DIFF has bought nothing. fix_arity_callers --any-proto (382 edits to the fleet-shared engine_core.h) and cast_self_callers --apply --sync-decls
(main's 4 self-conflicts) both did exactly that in S70: every target compiled afterwards and none matched.
Revert such an edit rather than leave unverified shared-state risk standing for zero banks.