docs(phase-29): func_8014D820 close=9 is a 3-profile MEASURED floor; backlog updated 25->9

.run/backlog.jsonl

@@ -1734,3 +1734,4 @@
 {"ts": "2026-07-24 20:23:40", "addr": "0x80174cb0", "name": "func_80174CB0", "reach": 138, "klass": "regalloc-order", "nins": 123, "status": "failed", "closeness": null, "where_stuck": "none \u2014 MATCH (123 ins)", "best_draft": ".run/backlog_drafts/func_80174CB0.c", "binary": "ov_SC01_077", "source": "t6-recover", "residual": null, "passes_tried": null}
 {"ts": "2026-07-24 20:24:20", "addr": "0x80156670", "name": "func_80156670", "reach": 138, "klass": "regalloc-order", "nins": 83, "status": "failed", "closeness": null, "where_stuck": "none \u2014 MATCH (83 ins). param_2/param_3 as s32 (not u16) flips the $s5/$s6 tie-break", "best_draft": ".run/backlog_drafts/func_80156670.c", "binary": "ov_SC01_077", "source": "t6-recover", "residual": null, "passes_tried": null}
 {"ts": "2026-07-24 20:24:20", "addr": "0x80174cb0", "name": "func_80174CB0", "reach": 138, "klass": "regalloc-order", "nins": 123, "status": "failed", "closeness": null, "where_stuck": "none \u2014 MATCH (123 ins)", "best_draft": ".run/backlog_drafts/func_80174CB0.c", "binary": "ov_SC01_077", "source": "t6-recover", "residual": null, "passes_tried": null}
+{"ts": "2026-07-24 21:50:11", "addr": "0x8014D820", "name": "func_8014D820", "reach": 138, "klass": "schedule", "nins": 304, "status": "near-miss", "closeness": 9, "where_stuck": "SESSION-18: 25->9 (304/304, prologue byte-exact). Residual is ONE schedule permutation in idx 84-98 (pos/desc block): mine issues z0=ent->z at 87 and stores desc.x at 90 / pos.y at 98; target issues z0 at 91, stores pos.y at 89 / desc.x at 98. MEASURED FLOOR: all THREE permuter profiles flat from THIS seed (regalloc 10 cycles, schedule 8, cse 8). Reading exhausted: 11 source-shape attempts, and every arrangement moving the z0 load later costs +1 ins while the target does it free => the placement is scheduler-produced, not expressible in C. symcheck 12/12 clean. Banking pre-cleared: 32 local_type (T0, auto) + 3 data_decl (T0, reconcile_decls) + 1 self_decl_hdr = the SS30#2 macro-widen in engine_core.h => FLEET-SHARED T2, R22 MANDATORY. Cheap tiers are now genuinely exhausted => this is the justified Fable5 case (Drew's call).", "best_draft": ".run/giants/s18_func_8014D820_close9.c", "binary": null, "source": "ov_SC01_077", "residual": null, "passes_tried": null}

phase-ends/CURRENT_PHASE.md

@@ -3250,3 +3250,20 @@ conditional) · main-EXE/B9 + GLM/B6 + resident's 14 walls (P30) · behemoths B7
   pressure and cascades the frame 0x40→0x38 into every save offset (§27 frame-pressure lock) — so
   expect the fix to trade 56 mismatches for a different set until the frame is re-locked. Untouched
   this session; this is the next giant to read after `func_8014D820`.
+
+- **📐 2026-07-24 (SESSION-18) — `func_8014D820` close=9 is a **MEASURED FLOOR**, by the corrected
+  §66d-4 rule, not a guess.** All THREE weight profiles now come back flat **from the same close=9
+  seed**: `regalloc` (10 cycles, 0 gain) · `schedule` (8 cycles, 0) · `cse` (8 cycles, 0). Combined with
+  reading being exhausted (11 source-shape attempts, and the +1-instruction law that proves the
+  placement is scheduler-produced), **both cheap tiers are genuinely spent on this function** — which is
+  a very different claim from SESSION-17's "at its measured permuter floor" at close=25, because that
+  one rested on a single profile.
+  **Backlog updated** (`tools/backlog.py log`): `func_8014D820`, reach 138, nins 304, closeness **9**
+  (was 25), class `schedule`, best draft `.run/giants/s18_func_8014D820_close9.c`, with the full
+  diagnosis + the pre-cleared banking path recorded in `where_stuck`.
+  **⇒ THIS IS NOW THE JUSTIFIED FABLE5 CASE, and it is Drew's call** (he set "no Fable5 for now" at the
+  SESSION-17 close and asked this session to "see if we need Fable5 later"). The doctrine's condition is
+  met exactly: cheap-Opus reading stalled, the permuter stalled across all profiles, and the residual is
+  a *single named mechanism* (a 9-instruction schedule permutation in one 15-instruction window) rather
+  than a vague wall — the shape Fable5 has historically cracked (§45 flagship, §52). Value if solved:
+  **×138**. Cost control: one isolated agent on this one function, per the parallel-isolated doctrine.