BFM-decomp/docs/frontier-s39.md

Frontier snapshot — P30 S39 close (2026-08-05, ~01:00)

Derived fresh from corpus.stubs + the sigs (regenerate any time: .run/s39/frontier_snapshot.txt is the raw output). Reconciled against every other "what's left" view by make audit-frontier — see that gate's output for disagreements.

The whole open frontier

13,312 open stubs / 741,792 instructions across 140 binaries.

The single biggest target is now main

binary	open ins	open stubs
main	79,074	2,001
ov_SC02_005	18,266	381
ov_SC04_011	15,366	285
ov_SC06_033	13,302	147
ov_SC06_032	13,149	169
ov_SC03_028	13,031	203
ov_SC02_011	13,017	260
ov_SC06_029	12,942	178
ov_SC06_018	12,504	169
ov_SC06_024	12,243	170

main is 10.7% of all open instructions — and it only became visible at this size tonight: its Ghidra sig was 7 weeks stale and covered 1,525 of 2,002 stubs, so 477 of them (19,309 ins) were invisible to every metric and every target selector. That is the R36/R34 lesson repeating: work a consumer cannot see does not exist to planning.

main is also the least tractable per instruction — it is structurally barren (zero h_exact overlap with the overlays; "main is not free dedup" was checked two ways in S31/32 and both were negative), so it is the genuine ~490-tok/ins agent tail, not a mechanical harvest.

Overlay open stubs by size (main excluded)

band	stubs	ins
≤20	2,003	31,413
21–50	5,243	173,862
51–150	3,246	264,784
151–400	765	155,807
>400	54	36,852

The mass is the 51–150 band (264,784 ins) — too big for the free local ≤15 tier, well inside what the §17–§45 toolkit handles. The >400 tail is only 54 stubs / 36,852 ins.

What is measured, and what is still assumed

• Measured tonight: the free h_exact pool (§142) is now 66 instances / 1,061 ins, all macro-backed — the cheap mechanical vein is essentially spent at this address-identical tier.
• Measured tonight: stored-draft re-gate rates by population — 4/6 fresh, 1/12 general stored, 3/17 reverted-overlay, 8/35 draft-exemplar heads (docs/calibration.md).
• Still assumed: the frontier report's h_norm cluster pool (1,689 clusters / 326,261 ins at a claimed 2.7× multiplier). Its whale claim was 3/4 wrong (§142), so this number must be re-measured before a wave is scoped onto it — that is S5's first step, not its premise.
• Still open (denominator): the 39 un-onboarded type-1 modules and main's missing independent boundary oracle. Neither is in the 741,792 above, because nothing can see them yet.

---

Burn-down (T5 prep) — derived from the committed digest history (R33), not a hand-kept series

day	instr-weighted	Δ
2026-07-22	78.65%	—
2026-07-23	79.36%	+0.72
2026-07-24	79.91%	+0.54
2026-07-25	80.57%	+0.66
2026-07-26	81.70%	+1.13
2026-07-27	84.40%	+2.70
2026-07-28	86.56%	+2.16
2026-07-29	87.18%	+0.62
2026-07-30	88.20%	+1.02
2026-07-31	89.16%	+0.96
2026-08-01	92.41%	+3.25
2026-08-03	93.71%	+1.29
2026-08-04	94.50%	+0.79
2026-08-05	94.36%	−0.14

⚠️ READ THE LAST ROW CORRECTLY — it is NOT a regression, and T5 must not report it as one. The numerator ROSE on 2026-08-05 (12,419,002 → 12,419,169). The percentage fell because the denominator grew 13,141,652 → 13,160,961 (+19,309) when main's stale sig was regenerated and 477 previously-invisible stubs entered the count. Same shape as Phase 27's 68.9% → 67.0% when four hidden overlays were onboarded: an honest denominator moves the headline down.

This is precisely the failure mode S1e caught at the other end of the same session — a percentage compared across a changed basis, read as a movement in the work. The rule (§140): a committed number is a claim about a tree; if it cannot be recomputed from that tree it is not evidence. Here both numbers are recomputable and correct; what changed is what they are numbers OF.

For phase-over-phase continuity the digest also prints the main-excluding series (fleet EXCLUDING main), which is the apples-to-apples comparison against pre-2026-07-22 readings.