BFM-decomp/phase-ends/PhaseEnd_Phase29.md

PhaseEnd — Phase 29: The Family Campaign (cores + members) — the biggest gain in project history, and the engine honestly spent

Date: 2026-07-30 · Project Version: 1.28.0 · Phase Status: Complete (milestone confirmed by Drew, gate 2) · Generation: Gen2 (21st phase of the arc)

Gen2 phase 21 (8→…→28→29). Third phase off the roadmap (v1 §3 P29). Ran 25 sessions over 15 days (2026-07-16 → 2026-07-30), 552 commits (2e84b53b2..8c7395027) — the longest and highest-yield phase of the project. The granular per-task trail (every wave verdict, every probe, the session checkpoints v1–v9, the error ledgers) is preserved on-demand at phase-ends/logs/Phase29.md (R19 — NOT auto-loaded; ~805 KB, consult only when researching a mechanism). This file is the synthesis. Owner decisions (Drew): the 2026-07-16 charter (longer multi-session campaign, ROI-gated, Fable5 reserved for new wall classes); the SESSION-16 integration-recovery re-plan (probe-then-continue with kill gates); the 2026-07-30 close call + the roadmap-v2 re-baseline directive (plan-mode gate, Fable 5/Max); milestone confirmed 2026-07-30.

Build Log

The headline (all three metrics, the §1 contract): instr-weighted 68.9% → 87.5% (+18.6pp) · distinct-code 49.5% → 78.0% (+28.5pp; 53,528 → 69,816 unique fns, +16,288) · fn-count 83.94% → 92.00% (+8.06pp) · 140/140 byte-identical at every banked batch (R22; 14× in SESSION-25 alone) · 0 NON_MATCHING linked throughout (G4) · dedup 1840 → 1886 groups · tools-health RC=0. The main EXE entered the weighted denominators 2026-07-22 (the roadmap §1 metrics contract, implemented; pre/post deltas not perfectly comparable across that date).

Files created/changed and complete — do not recreate (durable output; the archived log has the exhaustive per-commit list):

Source (the campaign): thousands of banked members across src/ov_*/*.c + src/shared/engine_core.h (the DEFINE_ macro corpus) + config/dedup.us.yaml (1840 → 1886 groups). SESSION-25 alone banked 2,713 members (+100,572 ins, +1,620 unique fns) across T79–T98. The §20 broad type-lift landed (154 types, SESSION-14, R22 140/140).

New/extended tooling (committed; headline items):

• tools/diff_regions.py (T1) — remap-exactly-as-staged + compile-at-exemplar-opt + masked region diff + per-member classification (O0-FLAG/TEMPLATES/PLUMBING-ISO/REGALLOC/NO-TARGET). The tool that resolved the roadmap's swing number.
• tools/lift_types.py (S13/S14) — the scoped §20 type-lift executor (clean-type lift → the broad 154-type lift).
• tools/recover_integration.py extended into the tiered T0/T1 blast-radius recovery driver (T16) + the §65b per-overlay de-macroize path — measured 14/36 = 39% recovery on the stranded-draft population.
• family_sweep --hseq hardening across the campaign (carry-macros, --no-preclassify, staging fixes), gate_stage reconcile ladder as the integration spine (call-site casts), jtbl_family_bank carve waves, imm_map/--fix-def-sig (§54/§119 — a REPAIR, not a default), permuter targeting fix (T13A).

Docs / knowledge base (R16/R30/R31):

• docs/matching-cookbook.md §54–§121 — 68 new sections in one phase (the idiom bank's largest growth), incl. §61/§63 blast-radius, §65b de-macroize, §83 backlog-is-not-a-queue, §116 opt-level-is-a-property-of-the-FILE, §117 spell-the-symbol-from-the-sibling's-address, §118 ordinal immediates, §119 off-diagonal lever testing, §120 prove-your-patch-RAN, §121 synthesize-externs-for-macro-defined-callees.
• docs/roadmap-to-100.md → v2 (the owed re-baseline, written at close under Drew's plan-mode gate): P30–P33 rechartered on 2026-07-30 measured reality; buckets W1–W5/M/R/T/B/D; member_adapt + the family-adapt fine-tune VACATED; v1 preserved in git history.
• docs/decision-log.md — ~15 new R31 entries (the swing resolution; the fresh-exemplar-sweep finding; the LIVE-siblings lever; the fix_header_decl→gate_stage doctrine correction; the §20 lift arc; the integration-bottleneck audit; psxport parked).
• docs/calibration.md — the swing number RESOLVED section + the stranded-draft recovery measurement.
• Digests regenerated throughout: progress.fleet.md (the per-session burn-down substrate), family-hseq.md, worklist.md, backlog.md.
• phase-ends/CURRENT_PHASE.md → phase-ends/logs/Phase29.md (R19). This file.

Tools/packages installed: None — the existing toolchain + venv throughout. Ultracode waves + cheap-Opus/Haiku-tier agents per the calibration doctrine; Fable 5 used for the close-out synthesis + roadmap v2 (its charter slot — new-wall-class discovery — was otherwise unused: no new wall class appeared that the idiom bank couldn't handle). No Ghidra DB change requiring commit (R23 — the db.*.gbf churn is restart-noise; the one MCP arc, ov_SC06_018 prefetch, was read-only).

Verification results (literal) — the milestone proof:

• make clean && make extract-all && make check-all → 140 passed, 0 failed of 140 byte-identical, reproduced at every banked batch across 25 sessions (14× in SESSION-25).
• make tools-health → OK (exit 0): corpus(+resident) 0 PHANTOM + 0 TRUNCATED; audit-binaries 140/140 citizens (R36); audit-cdecl green; dedup-check 1886 validated / 0 failed; report(lint+dedup) green.
• make report → the headline numbers above; 0 NON_MATCHING in any default build (G4).
• The burn-down (T99, derived from the committed digest history per R33): daily instr series 68.9 → 87.5 with final-four-day decay +2.7 → +2.2 → +0.6 → +0.3pp/day — the engine's measured floor.
• T98 residue characterisation: 80 families / 960 members / 173 distinct (29 all-STRUCT refused at remap by design + 51 no-STRUCT failing the gate with every lever applied; last three blast sweeps returned 0).

Milestone achieved (confirmed by Drew, gate 2): Phase 29 drove the fleet +18.6pp instr / +28.5pp distinct / +8.06pp fn-count at 140/140 byte-identical throughout — the largest single-phase gain in the project — resolved the roadmap's swing number ((a) TOOLING: the "~3% h_seq ceiling" was an -O0 compile-flag artifact; REGALLOC 0 on the 106-member sample), closed every named blocker in its long crack-and-bank series (not one was a compiler wall — each was a declaration, a symbol spelling, a flag combination, or a file placement), and then closed itself on measured evidence: the mechanical family engine's residue is 173 distinct against 2,713 banked in the final session, with daily yield decayed to +0.3pp. The close-out delivered the owed roadmap v2 re-baseline (P30–P33 on 2026-07-30 reality).

Next: Phase 30 — The Recovery & Concentration Campaign (roadmap v2 §3 P30), opening with T0 frontier regen + instrument repair (reconcile the family_hseq↔progress count gap; fix gate_stage snapshot/restore + rtu_match error surfacing; purge the backlog ledger) before any wave is scoped (R35). Then T1 integration-recovery sweep, T2 the -O0 Arm-A splat wall, T3+ the standing crack-wave loop. Fresh session, plan mode, Max (Tier-1 Phase-Start).

Deviations

Item	Plan (charter / roadmap v1 P29)	Actual	Reason
member_adapt.py + the member-harvest track	THE P29 member engine ("member_adapt + warm-start grinder over drift members")	NEVER BUILT — VACATED	Task 1 dissolved its premise: the h_seq DIFF class was an -O0 compile-flag artifact + type-lift plumbing, not adaptable drift; the harvest ran through family_sweep --hseq/dedup/gate_stage instead and banked thousands
The family-adapt fine-tune (charter item 4, bounded experiment)	tune the local 7B on the adaptation task	NEVER RUN — VACATED with member_adapt	same dissolved premise; the W4-STRUCT residue (86 members) cannot justify a fine-tune
Fable5 slot	reserved for NEW wall classes	unused for discovery; used at close for the roadmap-v2 synthesis (Drew-directed)	no new wall class appeared that cheap-Opus + the §31–§121 idiom bank couldn't handle — the doctrine held
Task 4 (grinder warm-start + 2 Phase-22 grinder bugs)	in-phase	CARRIED → P30	never load-bearing this phase (the wave machinery out-paced the grinder lane); genuinely useful only with P30's drift work
Task 5 (fleet Ghidra-C prefetch)	in-phase fuel	partial (ov_SC06_018); CARRIED → P31, re-scoped	SESSION-13 measured the gap is in the TAIL + main (the reach-138 head is 146/159 cached) — the prefetch pays off in P31, and needs MCP sessions (R23/R29)
Task 6 (permuter backlog sweep)	in-phase	partial (targeting fix T13A landed); sweep CARRIED → P30	fresh cracks + recovery out-yielded the permuter lane all phase; the close-1..4 drafts remain named fuel
Task 16 — integration-recovery pass	(not in the original plan)	ADDED mid-phase (Drew-approved plan, SESSION-16), measured, delivered	the SESSION-15 audit found the wave bottleneck is INTEGRATION (~92% of drafts byte-correct, ~27% bank) — recovering stranded byte-correct work beat drafting more
The -O0 ×138 rollout	implied by the Task-1 verdict ("the pool is back on the table")	WALLED — the Arm-A splat %lo +0x20 re-carve defect; CARRIED → P30 T2 as a named instrument task	the fix is splat-layer, not matching-layer; forcing it mid-campaign risked the fleet gate; the no-recarve driver alternative is sketched in the log
Roadmap maintenance	a PhaseEnd "Roadmap delta" line	full roadmap v2 rewrite at close	Drew's 2026-07-30 directive: 25 sessions of campaign had invalidated v1's numbers wholesale; a delta line couldn't carry the re-baseline
My errors (recorded, not buried — the SESSION-25 ledger + earlier)	(n/a)	§116 written with the wrong fix (refuted in 56s) · a bank predicted against a disqualifier I'd printed myself · "126 entries" that were 8 · T84's banks mis-attributed (committed wrong, corrected f7c6d2eb2) · T92's wrong recipe + phantom blocker (corrected 9a1507462) · the wrong staging site patched twice → a lever mis-judged from a patch that never ran · T97 costed as a 3-part 2,022-decl job that was one 4-line edit · and the month-old shared byte-gate defect found+fixed (SESSION-11: one binary verified against another's hash because a default was truthy)	the common root: asserting the composition of a number, or the effect of a change, I did not derive — distilled into §120 ("prove it RAN") and the R37 candidate below

Commit Message

(Phase 29 landed as 552 commits 2e84b53b2..8c7395027 + this PhaseEnd. Fleet 68.9 -> 87.5% instr /
49.5 -> 78.0% distinct / 83.94 -> 92.00% fn-count; 140/140 byte-identical throughout; 0 NON_MATCHING.
The granular trail -> logs/Phase29.md.)

feat(phase-29): the family campaign — +18.6pp instr in 25 sessions; engine spent on evidence; roadmap v2 (v1.28.0)

- T1 diff_regions.py RESOLVED the swing number: (a) TOOLING — the "~3% h_seq ceiling" was an -O0
  compile-flag artifact (REGALLOC 0/106); the third structural wall to resolve to our own tooling.
- The campaign: crack waves + propagate-behind-each-crack + the §20 broad type-lift (154 types) +
  the jtbl carve waves + the SC07-region sweeps. SESSION-25 alone: 2,713 members, +100,572 ins,
  +1,620 unique fns, every named blocker closed — NOT ONE a compiler wall.
- Doctrine shifts (decision-log): fresh cracks are the only lever that moves distinct-code; the
  propagation cap gates de-duplication, not coverage; the wave bottleneck is INTEGRATION (~92%
  byte-correct, ~27% bank) -> the tiered T0/T1 blast-radius recovery driver (14/36 = 39% measured);
  gate_stage call-site-casts are the integration spine (bulk header edits BREAK builds).
- T98 characterised the residue: 80 families / 960 members / 173 distinct (29 all-STRUCT by-design
  + 51 gate-failing, 3 sweeps at 0) -> T99 burn-down (derived from digest git history, R33):
  final-four-day decay +2.7 -> +0.3pp/day. Closed on the measurement, gate-2 confirmed.
- roadmap-to-100 v2: P30 Recovery & Concentration -> P31 Scope-Complete + Main & Resident ->
  P32 Behemoths+Walls -> P33 Verify+Flip. member_adapt + family-adapt fine-tune VACATED.
- Honesty ledger: 7 recorded errors + the month-old shared byte-gate default defect (found+fixed);
  the -O0 Arm-A splat wall named and carried as P30 T2 instrument work, never force-banked.
- cookbook §54-§121 (68 sections); ~15 decision-log entries; calibration re-measured; R37 proposed.

Rules Added This Phase

Rule

Reason

R37 — Probe before costing (proposed; Drew to accept/modify/reject). Before PRICING or SCOPING a job — a "three-part fix", a "126-entry backlog", a phase task — probe ONE member/instance to ground the estimate, derive any attribution/count from corpus.stubs (or the governing invariant) before asserting it, and diff the artifact to prove an edit actually ran before judging the lever it carried. Extends R14 (verify data vs bytes) and R35 (verify the instrument) to the estimating/attribution step — the one place they didn't yet bind.

The SESSION-25 error ledger's common root, seven times over: T97 was costed as a 3-part job over 2,022 decls and was ONE 4-line edit (read the symptom, never probed the failure); T84's 137 banks were attributed to the wrong family in a commit; a lever was judged ineffective from a patch that never ran (§120); "126 entries" were 8. Every instance was an asserted composition/effect that one probe would have grounded.

(No OTHER new governance rules — per the Phase-8+ precedent, the phase's techniques live in the cookbook (§54–§121) and its findings in decision-log.md/calibration.md/roadmap v2. The blast-radius taxonomy (T0 draft-only / T1 binary-local / T2 fleet-shared→R22-mandatory) is an engineering invariant, recorded in roadmap v2 §5 and enforced by the recovery driver, not a conduct rule. The shared byte-gate default defect (S11) is governed by the existing R35/R32; the fix_header_decl→gate_stage correction is R14 applied. G3/P9, R14, R16/R30, R22, R26/R27, R31, R32–R36 governed the phase and it reinforced all of them.)

PhaseEnd Changelog

v1.27.0 → v1.28.0 — Phase 29 complete (Gen2 phase 21; the family campaign). The project's biggest phase: 25 sessions, 552 commits, +18.6pp instr / +28.5pp distinct / +8.06pp fn-count at 140/140 byte-identical throughout. It opened by resolving the roadmap's swing number with a purpose-built instrument (diff_regions.py): the "~3% legacy h_seq ceiling" was (a) TOOLING — an -O0 compile-flag artifact, the third consecutive "structural wall" to dissolve into our own tooling — which put the legacy-PURE pool back on the table and vacated the planned member_adapt delta engine before a line of it was written. The campaign then ran crack-waves with propagation behind every crack, landed the §20 broad type-lift (154 types), and produced three doctrine shifts now encoded in roadmap v2: fresh cracks are the only lever that moves distinct-code; the propagation cap gates de-duplication, not coverage; and the wave bottleneck is INTEGRATION, not idioms (~92% of wave drafts byte-correct, ~27% banking) — answered with the tiered blast-radius recovery driver (measured 14/36 = 39%). SESSION-25 closed every named blocker on the carried list (2,713 members in one session), and not one was a compiler wall. The phase then did the rarest thing: it measured its own engine spent (T98: 173 distinct residue; T99: daily yield decayed +2.7→+0.3pp with every lever applied) and closed on the evidence, delivering the owed roadmap v2 (P30–P33 on 2026-07-30 reality). The honesty ledger records seven asserted-not-derived errors plus a month-old shared byte-gate default defect, found and fixed forward — the root distilled into the proposed R37 (probe before costing). Cookbook §54–§121; the -O0 Arm-A splat wall carried as named P30 instrument work.

Roadmap delta (roadmap §0 — what Phase 29 changed about the road ahead)

The delta IS roadmap v2 (docs/roadmap-to-100.md, committed 8c7395027 at close, Drew's plan-mode gate). Headlines: the family-remap engine is spent on measurement (W4 = 173 distinct); the three biggest un-spent levers are stranded-draft recovery (39% measured), the -O0 Arm-A splat instrument wall (~1,287 distinct), and the fresh-crack concentration head (top-100 ≈ 53% of remaining instr; 61 zero-crack families); main (0.7%), resident (14), and the 39 type-1 modules are the P31 flag-plant arc; behemoth ledger REBUILT (0x80183814, 5,122 ins, is the game's largest fn and was missing from v1); member_adapt + the family-adapt fine-tune VACATED; scale estimate P30 ≈ 10–18 sessions · P31 ≈ 4–8 · P32 unknown · P33 ≈ 2–4, with the flip-timing checkpoint live at every close.

Plain-English Recap

This phase was the big harvest. Over two weeks and 25 working sessions, the amount of the game rebuilt perfectly, byte-for-byte jumped from about 69% to 87.5% — the largest gain in the project's history — and all 140 pieces of the game still rebuild flawlessly after every single change.

Three things mattered beyond the number. First, the phase opened by settling an old question with a purpose-built measuring tool: last month's verdict that a whole class of "similar-but-shifted" functions could only be automated 3% of the time turned out to be our own build using the wrong compiler setting — the third time in a row a "wall" turned out to be our own equipment. Second, we learned where the real friction is: our AI drafters produce correct code ~92% of the time, but only about a quarter of it was making it into the final build because of paperwork — declarations, file placement, name spellings. So we built a recovery pipeline for stranded-but-correct work, and in the final marathon session every single stuck item on our list fell — 2,713 functions banked in one day, and not one blocker was the 1990s compiler; every one was paperwork. Third, when the machine finally ran out of cheap wins, we proved it with data rather than feelings — daily progress fell from +2.7 points to +0.3, and only 173 genuinely new functions remained in that engine's reach — and closed the phase on the evidence.

I also kept an honest list of my own mistakes this phase — seven times I asserted a number or an effect I hadn't actually checked, plus one verification defect that had sat unnoticed for a month — each is recorded, fixed, and distilled into a proposed new working rule: probe one example before pricing the job. The road ahead is now re-mapped on today's real numbers: recover the stranded correct code, fix one known extraction bug that's hiding ~1,300 functions, crack the biggest remaining function families, then finish the main program and plant the first 100% flags.

🛑 Stop Here

PhaseEnd written; CURRENT_PHASE.md archived → phase-ends/logs/Phase29.md (R19, via git mv — left uncommitted with this file for Drew's milestone-close commit, R6 gate-2 sign-off). The Phase-29 work is committed across the sessions (2e84b53b2..8c7395027, 552 commits incl. T99 + roadmap v2); Drew commits AND pushes this PhaseEnd + the archived log. No Ghidra DB staging (R23 — the db.*.gbf churn is restart-noise). The pre-existing .run/backlog.jsonl edit rides along or is discarded at Drew's discretion (it is a regenerable digest input). Gen2 continues — do NOT start Phase 30 here. Start a fresh session (effort Max, plan mode) for Phase 30 — The Recovery & Concentration Campaign (roadmap v2 §3), opening with T0 frontier regen + instrument repair before any wave is scoped. Keep this file forever.