FLEET 96.29% fn / 94.2% instr / 88.9% distinct, R22 140/140 twice. Session banked 7 wave-6 heads
+ 25 propagated siblings = +10,616 instructions (12,368,236 -> 12,378,852; the digest delta matches
the hand-derivation exactly).
wave-metrics.md gains Findings 5 and 6:
- RANK WAVES BY INSTRUCTIONS, NOT HEADS. Wave 6 banked ~45% more instructions than wave 5 while
banking less than half as many heads, because a bigger head carries more instructions AND its
family propagates at the same cost per sibling (func_8017FEE0: ONE 299-ins head -> 4,485 ins
across 15 siblings, ~0 agent tokens). The metric to beat is POOL REALISATION (21%), not bank
rate — and wave 6's bank rate is NOT comparable to waves 3-5 because the difficulty knob moved
deliberately (median target 438 ins vs 143, mostly has_mid_jr).
- A WAVE TALLY IS A COVERAGE CLAIM and needs its own assertion.
The 9 unbanked drafts are diagnosed and preserved, and 6 of them are ONE class worth cracking
first: `undefined reference` to a sibling that IS defined in the overlay but only via a
definition-side __asm__ alias (§37/§124) — the carve repartitions the object and separates the call
site from the alias definition. One fix frees 6 already-paid-for drafts and will recur in every
carve-heavy wave. The other two are genuine near-misses (close=47 REGALLOC-PERM, close=370).
Tooling committed for reuse: w6_pool.py (the pool derivation, now a script), w6_diag.py (runs the
REAL gate path — s36_capture.py splices without the carve and is wrong for jr targets),
w6_jtbl_prop.py (carve-path propagation with the per-family commit jtbl_family_bank requires).
Fleet 96.27 -> 96.28% fn-count / 94.1% instr / 88.6 -> 88.7% distinct (77,765
uniq). R22 clean-fleet: 140 passed, 0 failed of 140. dedup 1910/0.
16 targets / 16,884 templatable ins. 19 agents, 2.71M tokens, 82 min wall.
The gate banked 16/16 — the FIRST perfect gate of the session, and the first
needing NO reconcile at all. Sweep: +26 members / 3 failed across 17 overlays.
NEW: docs/wave-metrics.md — the wave-by-wave performance log, with the derivation
commands so future rows are COMPUTED, not hand-transcribed (R33). Four findings,
each recorded with its caveat rather than as a bare number:
1. THE PROMPT IS THE LEVER, AND THE AGENTS WRITE IT. Bank rate 76 -> 77 -> 100
-> 100 -> 100% with models and gate held constant. The jump was STEP 0 (a
magic-literal grep of src/, ahead of engine_core.h) — which came from a
wave-2 agent's index_gap report. Caveat recorded: waves 3-5 targets also
trended easier, so the mechanism is the durable claim, not the exact %.
2. pipeline() vs batched parallel(): 136 min/14 targets -> 82 min/16 targets,
parallelism 2.5x -> 3.8x. The two-batch design was a hard barrier with 37-50
min dead gaps; the harness already caps at 16 so it bought nothing. Floor
recorded honestly: the slowest agent is still ~50 min of real match_one
iteration, so the lever there is target SELECTION, not concurrency.
3. ECONOMICS: ~170-300k tokens per banked head in the stable regime — but a head
is not the unit of value. Head + propagated members is, and sweep yield is
BIMODAL not average (21/21 vs 18/165), because it is a property of the FAMILY.
Averaging those two predicts nothing.
4. A perfect gate is a signal the prompt rules landed. Waves 1-4 each needed 1-2
post-gate reconciles; wave 5 needed zero. The reconcile lane is the fallback,
not the plan. Lifetime 21/22.
Drew T