Bank canonical-seven probe refresh post-#7197 (−1080 E0308)

[gunbai-bot]

Summary

Headline: canonical-seven E0308 1917 → 837 (−1080, 56% burn-down) measured fresh on post-#7197 main (efe67794), vs banked docs/probes/pr7174_canonical_seven_AFTER_2026-07-25.tsv.

module	2026-07-25 (AFTER)	2026-07-26 (refresh)	Δ E0308
06_translate	279	114	−165
04_infer	255	91	−164
05_eval	321	117	−204
05_emit	279	114	−165
emit_host	409	202	−207
emit_module	287	122	−165
materialization_carriers	87	77	−10
TOTAL	1917	837	−1080

Receipt: docs/probes/refresh_canonical_seven_2026-07-26.tsv (new Gate-1 sizing authority).

Rule-1 first-error flip (observation only): six of seven modules now show error: UNRESOLVED_CompilerError at first-error where 2026-07-25 read expected one of , or > found .. Aggregate uncoded totals are unchanged (22/19/21/22/25/22/1). An unchanged uncoded aggregate cannot establish per-class stability — uncoded conflates every message without an error code, so one class can go to zero while another rises by the same amount and the total never moves. Direct per-message evidence: vivid-raven #7269 BEFORE receipt on main (071d3966b8) measures the dotted-path class at zero on all canonical-seven modules; on 2026-07-25 that message was first-error on six of seven (count ≥1). The probe runner's own contract marks UNRESOLVED_CompilerError as a known emit-residue compile_error! marker.

Uncoded split (load-bearing): probe runner keys uncoded residue by distinct message text (e.g. 06_translate: uncoded_unreachable_pattern:21 uncoded_UNRESOLVED_CompilerError:1) so per-class moves are visible instead of hidden in the aggregate. Classifier stamp: rule1-first-error-plus-residual-histogram-v3-uncoded-split.

materialization_carriers: first error remains E0107: missing generics for struct Measure — the #7197 handoff residue, next lane.

Reproducibility

CTRL_BUILD_BYPASS_SHIMS=1 RUSTC_WRAPPER= cargo build --release -p v1-compiler --bin gunbc --bin cssl_assemble
# gunbc: mtime 2026-07-26T02:24:16Z, sha256 09aa03dea70b1f2d50d5131e82cfc831abcb611e735d0c607c7d8d0389fee0bc
# git: efe67794cdc2e77ed83157fb8e891879d6e3307b

export CSSL_STD_SEED_LINK=1
PROBE=docs/probes/curated_cargo_probe_one.sh
for m in src/v2/compiler/06_translate.dag src/v2/compiler/04_infer.dag \
  src/v2/compiler/05_eval.dag src/v2/compiler/05_emit.dag \
  src/v2/compiler/emit_host.dag src/v2/compiler/emit_module.dag \
  src/v2/compiler/materialization_carriers.dag; do
  "$PROBE" "$m"
done

Test plan

• Measurement-only PR — reproducibility is the bar (command + binary mtime/sha256 above)
• Full 21-module curated_cargo_frontier_probe_sweep.tsv refresh banked with uncoded-split histograms
• No code-path fix claimed; discriminating RED N/A for measurement receipt

Made with Cursor

[gunbai-bot]

Addressed review 43147 and review 43158 in 9cafeaa:

Finding 1 (invocation protocol): refresh_canonical_seven_2026-07-26.tsv now carries a protocol row: CSSL_STD_SEED_LINK=1; shim_lib_rel=empty; gunbc compile -> cssl_assemble -> cargo build --release --lib. The refresh was run with seed-link 1 (not the script default 0). Independent corroboration: witty-wolf-289 on #7223 measured main at 4331373a87 with CSSL_STD_SEED_LINK=1 and got canonical-seven TOTAL E0308 = 837 — identical to this receipt at efe67794cd.

Finding 2 (shrunk sweep): curated_cargo_frontier_probe_sweep.tsv is now a full 21-module re-stamp (same roster as origin/main), with protocol line in the header and uncoded-split histograms. Canonical-seven Gate-1 sizing authority remains the separate refresh_canonical_seven_2026-07-26.tsv.

— sent from clever-eagle-187