mirror of
https://github.com/Druthulu/BFM-decomp
synced 2026-09-28 06:49:47 -04:00
feat(tells): land the crib, the grep aliases and per-lane budgets
Three measured harness defects, all fixed: 1. THE CARD NAMED A WORD THE COOKBOOK DOES NOT CONTAIN. api_agent stated the lever as a bare label and nothing else; grep 'extend-tell' / 'swaprepeat' / 's16-div-tell' over the 750-section cookbook returns ZERO. 108 failure transcripts grepped extend-tell and 28 grepped swaprepeat against nothing while the knowledge sat at 172a/172b under different words. Fixed both ends: a LANE ALIASES grep-bait block at 172b, and LEVER_CRIB on the card — what the tell means, the section to grep, and the byte-proven C spellings. 2. ONE GLOBAL AGENT BUDGET FOR CARDS OF VERY DIFFERENT SIZE. tells cards are 2.4x the default lane's (median 89-95 ins vs 37-39) and stack 3-5 idioms; 98 of 270 final attempts ended AT the 24-turn cap. LANE_BUDGET gives tells 40 turns / /bin/bash.40, and logs the choice so it is auditable rather than invisible. 3. Two new SYS laws: grep the section your crib cites before drafting, and stop when the residual class says [permuter]; plus 263 (an invented argument changes scheduling — check arity before reaching for a fence the permuter cannot help with). Cookbook 264 records the four recipes the tells agent drove to MATCH: the inline (s16) in a call argument, the save-order/bb0 anti-dependence law (new), the opaque-bound local assigned late (new), and 172b-1's multi-def mirror variable made concrete.
This commit is contained in:
Drew T
+589
-587
File diff suppressed because it is too large
Load Diff
@@ -16572,6 +16572,16 @@ repeated operand expressions — that redundancy is load-bearing for byte-matchi
|
||||
|
||||
## §172b — THREE MORE TELLS FROM THE GCC READ (P30 S50-Max, banked on Drew's ask)
|
||||
|
||||
**LANE ALIASES (grep bait — added P31 S59).** The atlas lever labels a drafting card carries
|
||||
(`tools/atlas.py`) do not appear anywhere else in this cookbook, so an agent that greps the label it
|
||||
was handed finds NOTHING: measured over the tells waves, **108 failure transcripts grepped
|
||||
`extend-tell` and 28 grepped `swaprepeat` against zero hits** while the knowledge sat here under
|
||||
different words. The map:
|
||||
`extend-tell` → **§172b-1** (+ the C recipes in §264) · `swaprepeat-tell` → **§172a** (macro-vs-inline
|
||||
repeats) + **§172b-2** · `s16-div-tell` → **§172b-4**. If you arrived by grepping a card label, this
|
||||
is your section.
|
||||
|
||||
|
||||
**1. The sll/sra-16 pair tell (extendhisi2 is MEM-only).** The single-insn sign-extending load
|
||||
(`extendhisi2_internal`) exists ONLY for memory operands. A visible `sll X,Y,16; sra X,X,16` pair
|
||||
in the target is therefore a promotion of a REGISTER-held short — a multi-def s16 variable (a
|
||||
@@ -24869,3 +24879,50 @@ levers and twenty CPU-minutes went to a defect one deleted argument closed.
|
||||
function takes the struct pointer, or because a seed's sibling had it. `decl_prior` (the fleet
|
||||
declaration consensus on the card) answers this when the symbol is known fleet-wide — read it before
|
||||
inventing a signature.
|
||||
|
||||
## §264 — FOUR TELLS-LANE C RECIPES, EACH DRIVEN TO MATCH (P31 S59)
|
||||
|
||||
The `extend-tell` / `swaprepeat-tell` / `s16-div-tell` population banks well below the default lane,
|
||||
and the measured cause is not the models: **the card names a lever that appears nowhere in this
|
||||
cookbook** (see the LANE ALIASES block at §172b) and the drafts stack 3-5 idioms. These four were
|
||||
proven this session by taking real failed wave drafts to `MATCH`; full derivations, transcripts and
|
||||
probe files in `docs/tool-designs/tells-lane-s59.md` and `.run/s59_tells/`.
|
||||
|
||||
**1. The `(s16)` promotion belongs INLINE in the call argument** (`func_8017D8E8`, 33/33). Spelled
|
||||
inline, its `sra` lands in the `jal` delay slot; hoisting it into a temp kills that fill. Three
|
||||
companions in the same function: hold the table base in a *source pointer local* so it owns its
|
||||
`$sN` across calls (gcc-2.7.2 has no cross-bb CSE, L2425); index with `p[i]` so `(field<<3)+base`
|
||||
stays ONE association unit (writing `base + f8 + i2` lets fold re-associate the constant out); read
|
||||
an s16 field directly in the expression for a plain `lh`. The failing draft had pinned EIGHT
|
||||
registers to transcribe the asm 1:1 and plateaued at near-2 over 23 oracle calls — **pins are the
|
||||
last resort (§17), not the opening move; they block the parm-copy and arg-copy shapes.**
|
||||
|
||||
**2. NEW LAW — a save-order-only residual is a bb0 anti-dependence** (`func_801808AC`, 67/67).
|
||||
Symptom: your prologue emits `sw $s1` before `sw $ra` where the target has `$ra` first, and nothing
|
||||
else differs. Saves are emitted descending-regno and woven at sched2, so an anti-dependence from a
|
||||
**basic-block-zero write of that `$sN`** promotes its store. Cure: move the first write of that
|
||||
variable out of bb0 and into a branch arm — `i = 0; if (…) goto found; i = 1;` becomes
|
||||
`if (…) i = 0; else if (…) i = 1; … else return;`. reorg then re-creates the target's
|
||||
`addu $s1,$zero` in the `beq` delay slot by thread fill.
|
||||
|
||||
**3. NEW LAW — an opaque bound must live in a LOCAL, assigned LATE** (`func_80180E58`, 34/34).
|
||||
`if (v < 0x10000)` and `if (v <= 0xFFFF)` both canonicalize to `li 0xFFFF; slt(0xFFFF,v); bnez`.
|
||||
The target's `lui 0x10000; slt(v,K); beqz` — plus the arm's store REUSING that register — is only
|
||||
reachable with the bound in a local, and assigned **late**, so its live range starts after the block
|
||||
temps die (assigned at the top it takes `$a1`, not `$v1`). Two companions: use the PARAMETER as the
|
||||
multi-use variable (its expand-time parm copy survives exactly when `$aN` is re-used; a
|
||||
`self = arg0` local emits double moves), and name the inner term of a nested `a - (b - K)` as a temp
|
||||
or fold distributes it away.
|
||||
|
||||
**4. §172b-1 made concrete — a naked `sll/sra` pair on a REGISTER is a branch-merged multi-def
|
||||
variable** (`func_8018A6A0`). Not a cast of a fresh load: mirror the memory slot with a variable
|
||||
written on BOTH paths (`vec0[1] = w = *s1;` at entry, `w = vec0[1] + …; vec0[1] = w;` in the loop)
|
||||
and compare `(s16)w` after the join. `(s16)vec0[1]` instead emits a fresh `lh` — store-forwarding
|
||||
cannot cross a two-predecessor label. This flipped the residual from structural to a 3-instruction
|
||||
commutative register tie, i.e. from an idiom problem to permuter fuel.
|
||||
|
||||
**Two honest walls recorded so nobody re-spends them:** `func_801AE3A4`'s missing same-value copy is
|
||||
reachable by an `"0"(b)` re-tie but sits behind §172's +8-frame canonicalization wall; and the
|
||||
`swaprepeat` siblings `func_80187E08`/`func_80182FD0` already carry the right doubled-compare
|
||||
structure — their 2-10-instruction scratch permutation is the S11 knife-edge, permuter fuel, and one
|
||||
fix banks both by remap.
|
||||
|
||||
@@ -0,0 +1,305 @@
|
||||
# Tells lane S59 — measured taxonomy, byte-proven idioms, and the path to 75%
|
||||
|
||||
**Investigation date:** 2026-08-24 (P31 S59). All artifacts under `.run/s59_tells/` (inventory,
|
||||
re-run results, streams, gap scans, and every probe `.c`). Nothing in `tools/` or `src/` was
|
||||
edited; every change below is a proposal.
|
||||
|
||||
**The one-line answer:** the tells lane is not failing on the tells. It is failing on (1) a
|
||||
**discovery gap** — the card vocabulary (`extend-tell`) has zero hits in the cookbook, so 650
|
||||
transcripts grepped and missed; (2) **stacked micro-idioms** — a tells function carries 3–5
|
||||
independent gcc-2.7.2 spelling laws, and missing any one reads as LENGTH-DRIFT; (3) **pipeline
|
||||
losses** — 26% of in-band failures are drafts that already shape-match and reloc-agree, lost at
|
||||
gate/integration, and another 12% are permuter-bucket residuals no permuter ever visits; and
|
||||
(4) **budget caps** — 36% of final attempts ended at the turn/cost cap, not at surrender.
|
||||
5–50 ins is cumulatively at 75% already; 50–80 needs only the pipeline fixes; 80+ is where the
|
||||
real wall is, and 75% there is NOT supported by current evidence.
|
||||
|
||||
---
|
||||
|
||||
## 1. Denominators (assert before believing any rate below)
|
||||
|
||||
Completed tells-lane waves at analysis time: **ao, as, at, au, aw, az, bb, bd, bg, bi** (wave bk
|
||||
was in flight and is excluded; `at` was a mixed draw that also carried non-tell levers).
|
||||
Cards → `.run/wave_??_cards.json`; drafts → `.run/wave_??/shard*/`.
|
||||
|
||||
| population | count |
|
||||
|---|---|
|
||||
| unique (binary,fn) carded across those 10 waves | 1,120 |
|
||||
| … of which carry a tell lever (extend 406 / swaprepeat 46 / s16-div 3) | 455 |
|
||||
| banked by now (INCLUDE_ASM gone from the card's TU, any wave/retry) | 730 |
|
||||
| still open | 390 |
|
||||
| still open **with a draft on disk** (newest per fn) | 270 (29,784 ins, median 91) |
|
||||
| still open with **no draft ever** (worker exhaustion; 43 UNKNOWN / 38 head-crack / 28 extend-tell) | 120 (10,437 ins) |
|
||||
| of the 270 re-run through `match_one --json`: | **near 216 · shape-MATCH 45 · toolfail 9** |
|
||||
|
||||
Re-run harness: `.run/s59_tells/rerun.py` (12-way, `--work` private per binary). "Banked" was
|
||||
derived from the live tree (a campaign is banking concurrently, so a function open NOW definitely
|
||||
failed its waves; one banked later by another lane counts as banked — the right population for
|
||||
"what does the lane still need").
|
||||
|
||||
## 2. The residual taxonomy, measured
|
||||
|
||||
`tools/residual_class.py` verdicts over the 216 nears (count / instructions at stake):
|
||||
|
||||
| klass | extend-tell | swaprepeat | s16-div | non-tell riders | share of nears |
|
||||
|---|---|---|---|---|---|
|
||||
| LENGTH-DRIFT | 105 / 12,901 | 11 / 1,577 | 1 / 110 | 23 / 3,465 | **65%** |
|
||||
| OPCODE-MIXED | 20 / 2,014 | 1 / 102 | — | 7 / 766 | 13% |
|
||||
| SIZE-MISMATCH (redraft: wrong body) | 5 / 511 | 2 / 292 | — | 4 / 505 | 5% |
|
||||
| SCHEDULE-REORDER | 6 / 496 | 1 / 76 | — | 3 / 225 | 5% |
|
||||
| WIDTH | 2 / 178 | 1 / 51 | — | 5 / 388 | 4% |
|
||||
| REGALLOC-PERM | 4 / 371 | 2 / 204 | — | — | 3% |
|
||||
| ADDRESSING | 5 / 559 | — | — | 1 / 45 | 3% |
|
||||
| other (REGALLOC-LOCAL, IMM-VALUE, BRANCH-POLARITY, DELAY-SLOT, STRENGTH) | 4 | 1 | — | 3 | 4% |
|
||||
|
||||
Buckets: **permuter 21** (16 of them at closeness ≤ 8), **structural 184**, **redraft 11**.
|
||||
|
||||
**LENGTH-DRIFT decomposed** (`.run/s59_tells/ld_scan.py`, greedy resync alignment over the
|
||||
`--emit-streams` word pairs, 140/140 scanned): gap mass ≤ 3 words: **31** · 4–8 words: **55** ·
|
||||
> 8 words: **54**. So under half of the dominant class is "one construct spelled wrong"; the rest
|
||||
are wrong *regions* — misread bodies, which no cookbook section fixes.
|
||||
|
||||
**Construct-delta measurement** (170 tell-lever nears, draft-minus-target counts,
|
||||
`.run/s59_tells/analyze3.py`): drafts run SHORT (Σdlen −417; 78 short vs 46 long), with fewer
|
||||
`lh` (−79), fewer `sll` (−64), smaller frames (Σ −288 bytes, 39 drafts smaller vs 15 larger), and
|
||||
MORE nops (+60). **`d_extpair = 0` for 74% of failing extend-tell drafts** — the sll/sra pair
|
||||
count is usually already right. The standing mental model "tells drafts miss the extend pairs" is
|
||||
refuted; what they miss is *placement* (which side of a branch/jal, fold-vs-materialize) and the
|
||||
surrounding fold/alloc laws. The bodies are systematically too *clean*: the original source's
|
||||
redundancy (textual repeats, mirror variables, opaque bound locals) is load-bearing and the
|
||||
drafts don't write it.
|
||||
|
||||
## 3. The discovery gap (why the cookbook isn't reaching the drafts)
|
||||
|
||||
From the 796 transcripts behind still-open fns (`.run/s59_tells/mine_transcripts.py`):
|
||||
650 grepped the cookbook, 524 read sections — the harness works. But:
|
||||
|
||||
- **`grep 'extend-tell'` — 108 times. `grep 'swaprepeat'` — 28 times. Both have ZERO hits** in
|
||||
`docs/matching-cookbook.md` (verified). `s16-div-tell`: zero. `grep '172'`: **zero** attempts.
|
||||
- The knowledge exists at §172a (L16557) / §172b (L16573) under *different words*
|
||||
("sll/sra-16 pair tell", "swapped-arm select tell", "extendhisi2 is MEM-only") — invisible to a
|
||||
model searching the only vocabulary it was given.
|
||||
- The card's entire statement of the lever is `tools/api_agent.py:384-385`:
|
||||
`"The atlas labels this card's lever: extend-tell."` — no meaning, no §, no recipe.
|
||||
|
||||
The agents' own submit notes (1,101 collected, 420 deduped, 274 non-MATCH,
|
||||
`.run/s59_tells/submit_notes.json` / `near_notes.json`) repeatedly *re-derive* §172b from the
|
||||
bytes ("target splits the counter's sign-extend as lhu; sll [delay slot]; sra; slt — every
|
||||
spelling I reached folds to lh", ov_SC02_005 func_8018A6A0) and then plateau without the recipe.
|
||||
|
||||
## 4. Byte-proven idioms (each verified with `match_one` this session; probes in `.run/s59_tells/`)
|
||||
|
||||
The whole-binary gate remains the arbiter — these are relocation-masked standalone MATCHes with
|
||||
the symbol walk done by hand against each target's own reloc lines (law 1c).
|
||||
|
||||
### 4.1 `func_8017D8E8` (ov_SC03_108, 33 ins) — near-2 pin-transcription → **MATCH**
|
||||
Draft `.run/wave_bg/shard114/func_8017D8E8.c` pinned EIGHT registers (incl. caller-saved
|
||||
$2/$3/$4-$6) to transcribe the asm 1:1 and plateaued at near-2 over 23 oracle calls. Natural C
|
||||
matched (`.run/s59_tells/fix_8017D8E8.c`):
|
||||
```c
|
||||
void func_8017D8E8(s32 arg0) {
|
||||
s32 i; u16 *tbl;
|
||||
for (i = 0, tbl = D_8018476C; i < 4; i++)
|
||||
func_8012C658(0x199,
|
||||
(s16)((i << 10) + *((u16 *)((*(s16 *)(arg0 + 0x70) << 3) + (s32)tbl) + i)),
|
||||
arg0);
|
||||
}
|
||||
```
|
||||
Levers, in order of necessity: **(a)** the `(s16)` promotion spelled INLINE in the call argument —
|
||||
its `sra` lands in the jal delay slot (a temp kills this); **(b)** the table base as a *source
|
||||
pointer local* so it owns `$s2` across calls (cookbook L2425 rule 1 — gcc-2.7.2 has no cross-bb
|
||||
CSE); **(c)** `p[i]` indexing so `(field<<3)+base` stays one association unit (writing
|
||||
`base + f8 + i2` lets fold re-associate the constant out); **(d)** comma-init order `i` then
|
||||
`tbl`; **(e)** `lh` = read the s16 field directly in the expression (§172a).
|
||||
|
||||
### 4.2 `func_801808AC` (ov_SC05_018, 67 ins) — prologue save-order swap → **MATCH**
|
||||
Draft `.run/wave_bg/shard117/…` plateaued at near-2: `sw $s1` emitted before `sw $ra` (target:
|
||||
$ra first). 43 match_one calls; its cookbook greps were "frame size", "delay slot|reorg" — wrong
|
||||
vocabulary. Root cause (sched.md S7 + `sched.c`): saves are emitted descending-regno and woven at
|
||||
sched2; an anti-dependence from a **bb0 write of $s1** promotes `sw $s1`. The draft's
|
||||
`i = 0; if (tbl == &A) goto found; i = 1; …` places `i = 0` in bb0. The else-if spelling moves
|
||||
every `i = K` into its own arm, bb0 stays clean, the fixed order survives:
|
||||
```c
|
||||
if (tbl == &D_801B3DC4) i = 0;
|
||||
else if (tbl == &D_801BBAB8) i = 1;
|
||||
else if (tbl == &D_801C29C4) i = 2;
|
||||
else if (tbl == &D_801CB724) i = 3;
|
||||
else return;
|
||||
```
|
||||
(`.run/s59_tells/fix_801808AC.c`, MATCH 67/67.) **General law, new to the cookbook:** a
|
||||
save-order-only residual ⇒ find which callee-saved reg is written in bb0 and move that first
|
||||
write into a branch arm (set-in-arm vs set-then-test). reorg then re-creates the target's
|
||||
`addu $s1,$zero` in the beq delay slot by thread fill.
|
||||
|
||||
### 4.3 `func_80180E58` (ov_SC06_000, 34 ins) — LENGTH-DRIFT/−1 stack → **MATCH**
|
||||
Draft (near-20, pinned $5/$6/$7) vs natural C (`.run/s59_tells/fix_80180E58.c`):
|
||||
```c
|
||||
void func_80180E58(u8 *self) { /* (a) param IS the variable */
|
||||
s32 amt = 0xFFFE4000; u16 *p = (u16 *)&D_801274EA;
|
||||
s32 lim; s32 t; s32 d; s32 v;
|
||||
d = *(u16 *)(self + 0xA) - 0x300; /* (b) fold-blocker temp */
|
||||
t = *p - d;
|
||||
v = *(s32 *)(self + 0x204) + amt;
|
||||
*(s32 *)(self + 0x204) = v;
|
||||
lim = 0x10000; /* (c) opaque bound, late */
|
||||
if (v < lim) *(s32 *)(self + 0x204) = lim;
|
||||
*p = *p - *(u16 *)(self + 0x206);
|
||||
if ((s16)t < 0) { *(s32 *)(self + 0x200) = 0x1E; func_80171990(self); }
|
||||
}
|
||||
```
|
||||
**(a)** Use the *parameter* as the multi-use variable — its expand-time parm copy
|
||||
(`addu $a2,$a0`) survives exactly when `$a0` is re-used (here by `t` and the call's own arg
|
||||
copy); a separate `self = arg0` local emits double moves, and a `$6` pin blocks `t` from taking
|
||||
`$a0` and lets cse pass the raw param to the call (killing the target's `addu $a0,$a2` slot
|
||||
fill). **(b)** `a - (b - K)` gets distributed by fold; a named temp for the inner term keeps the
|
||||
target's `addiu −0x300; subu` shape. **(c)** `if (v < 0x10000)` and `if (v <= 0xFFFF)` BOTH
|
||||
canonicalize to `li 0xFFFF; slt(0xFFFF,v); bnez` — the target's `lui 0x10000; slt(v,K); beqz`
|
||||
plus the store *reusing the same register* is only reachable with the bound in a local; assign it
|
||||
LATE so its live range starts after the block temps die (assigned at the top it allocates $a1,
|
||||
not $v1). **(d)** `(s16)t < 0` emits `sll 16; bgez` (sra elided) for free.
|
||||
|
||||
### 4.4 `func_8018A6A0` (ov_SC02_005, 76 ins) — §172b-1's C recipe made concrete
|
||||
Draft's compare `(vec0[1] << 15 << 1)` → structural STRENGTH/lhu!=sll. `(s16)vec0[1]` → a fresh
|
||||
`lh` (the compare sits after a two-predecessor label; store-forwarding cannot cross it). The
|
||||
target's naked `sll/sra` pair on a REGISTER is a **branch-merged multi-def variable** — §172b-1's
|
||||
"like `mw`", spelled:
|
||||
```c
|
||||
vec0[1] = w = *s1; /* entry def */
|
||||
…
|
||||
w = vec0[1] + *(u16 *)(s1 + 2); /* loop def: lhu (real mem re-read) + addu */
|
||||
vec0[1] = w;
|
||||
check:
|
||||
if ((s16)w < *(s16 *)(s1 + 1)) goto body; /* sll/sra of w after the join */
|
||||
```
|
||||
This flipped the residual from structural to REGALLOC-PERM/3 (a pure $v0/$v1 commutative-tie,
|
||||
`[permuter]` bucket) — the promotion bytes now exact. Six tie spellings probed
|
||||
(`.run/s59_tells/v_*.c`), all stable at 3: that last mile is permuter fuel, not an idiom.
|
||||
|
||||
### 4.5 The honest walls met
|
||||
- `func_801AE3A4` (25 ins, LD/−1): the missing `addu $v1,$v0` is a same-value copy that plain
|
||||
`b = a` cannot produce (cse deletes it; a `"0"(b)` re-tie materializes it — mechanism confirmed
|
||||
— but the target's +8 frame slot is §172's documented three-layer canonicalization wall).
|
||||
Left unsolved; routed to §172's instrument, not to more spelling probes.
|
||||
- `func_80187E08` / `func_80182FD0` (swaprepeat siblings, 102 ins): the doubled compare structure
|
||||
(§172b-2) is already RIGHT in the drafts; the residual is a 2–10-ins scratch permutation that
|
||||
flips between spellings (S11 knife-edge). Permuter fuel; one fix banks both via remap.
|
||||
|
||||
## 5. Pipeline losses (free wins, no model change)
|
||||
|
||||
1. **45 shape-MATCH drafts sit unbanked** (3,105 ins) — 34 of them also reloc-**AGREE**
|
||||
(`.run/s59_tells/shape_matches.json`). 9 are on `main`/md_MAIN, where wave gating cannot bank
|
||||
(gate_main is a separate, slate-level path). The remaining ~25 overlay fns died at the
|
||||
whole-binary gate on TU plumbing. §173's measured recovery rate on exactly this class is 64%
|
||||
(`aprop_symfix --fix` → `recover_integration --stages macro-externs,demacroize,tu-scope`).
|
||||
2. **21 permuter-bucket nears** (16 at closeness ≤ 8, 1,710 ins) — the lane never runs the
|
||||
permuter. `residual_class` already routes them (`match_one --json` carries the verdict); a
|
||||
zero-token post-wave permuter pass over `bucket=="permuter" && closeness<=8` drafts is the
|
||||
designed use of `tools/permuter/`.
|
||||
3. **Budget caps bind**: of the 270 newest attempts, 98 ended at the turn cap and 12 at
|
||||
cost/API stops (`submit`ted: 160). `tools/ox_campaign.py:264` pins every lane to
|
||||
`--max-turns 24 --max-cost-per-fn 0.15`; tells cards are 2.4× default size (median 89–95 vs
|
||||
37–39) and my three hand-cracks each took 5–10 oracle iterations WITH the right hypothesis
|
||||
already in hand. Wave ao alone logged 562 truncated turns for 185 targets.
|
||||
4. **120 drawn cards never got a draft at all** (10,437 ins) — worker exhaustion in the 2,000-way
|
||||
fan-out; they re-enter only via `--retry-unbanked`.
|
||||
|
||||
## 6. Concrete changes (proposed; nothing landed)
|
||||
|
||||
### 6.1 Cookbook — make the card vocabulary greppable (3 lines, highest ROI in this doc)
|
||||
At the §172b header (`docs/matching-cookbook.md:16573`) append an alias line, and mirror it in
|
||||
`docs/cookbook-index.md:1088`:
|
||||
|
||||
> **Lane aliases (grep bait):** `extend-tell` = §172b-1 (+§4.1–4.4 recipes below), `swaprepeat-tell`
|
||||
> = §172a macro-vs-inline + §172b-2, `s16-div-tell` = §172b-4. Atlas lever labels from
|
||||
> `tools/atlas.py:441-445` — a drafter grepping the label must land here.
|
||||
|
||||
Then add the four byte-proven laws of §4 above as proper cookbook sections (the save-order/bb0
|
||||
anti-dep law of 4.2 and the opaque-bound-local law of 4.3 are NEW; 4.1 and 4.4 sharpen §172b-1
|
||||
into C recipes). Per R30/bank-idioms-before-checkpoint, this belongs in the same session that
|
||||
lands the doc.
|
||||
|
||||
### 6.2 `tools/api_agent.py` — put the crib ON the card (replace lines 384-385)
|
||||
```python
|
||||
LEVER_CRIB = {
|
||||
'extend-tell': (
|
||||
"extend-tell = §172b-1 (grep 'sll/sra-16 pair tell'). The target promotes REGISTER-held\n"
|
||||
"s16 values. The four placements and their C spellings:\n"
|
||||
" * naked sll/sra pair = a MULTI-DEF s16/u16 variable merged across branches — write\n"
|
||||
" `w` assigned on BOTH paths (mirror the mem slot: `x[i] = w = …`), compare (s16)w;\n"
|
||||
" * plain lh = single promoted use — read the s16 mem DIRECTLY in the expression, no temp;\n"
|
||||
" * sll before a jal + sra IN its delay slot = the (s16) cast INLINE in the call argument;\n"
|
||||
" * lhu + separate sll/sra = the value is also used as u16 — keep the u16 variable.\n"
|
||||
"Also: sign test (s16)t<0 emits sll+bgez only. NO register pins first (§17 is last resort):\n"
|
||||
"pins block the parm-copy/arg-copy shapes. Param-as-variable: if a param is copied to $aN,\n"
|
||||
"type the PARAMETER and use it everywhere; do not invent `self = arg0`.\n"),
|
||||
'swaprepeat-tell': (
|
||||
"swaprepeat-tell = §172a/§172b-2 (grep 'swapped-arm select tell'). Textual REPEATS are\n"
|
||||
"load-bearing: a repeated compare/select re-emits its own slt+branch; write the repeat\n"
|
||||
"TEXTUALLY (nested macros repeat operands; inline fns collapse — §172a). Range checks are\n"
|
||||
"plain `if (-lim < x && x < lim)` — the 2nd slt rides the 1st beqz's delay slot free.\n"),
|
||||
's16-div-tell': (
|
||||
"s16-div-tell = §172b-4 (grep 'Division-by-constant sign-correction'). s16/K magic-multiply:\n"
|
||||
"keep the promotion as an EXPRESSION (not a temp) so the sign-sra can reuse the <<16\n"
|
||||
"intermediate; K's magic pairs are in §167-25 and the odd-part table.\n"),
|
||||
}
|
||||
if card.get('lever'):
|
||||
out.append(f"\nThe atlas labels this card's lever: {card['lever']}.")
|
||||
crib = LEVER_CRIB.get(card['lever'])
|
||||
if crib:
|
||||
out.append(crib)
|
||||
out.append("Universal fold-blockers when a shape refuses: name the inner term of a\n"
|
||||
"nested +/- as a temp (keeps association); hold a non-slti constant bound in\n"
|
||||
"a LOCAL assigned right before its `if` (defeats the 0xFFFF canonicalization\n"
|
||||
"and lets the arm's store reuse its register); hold a global base in a\n"
|
||||
"pointer local when it must live in an $sN across calls (L2425).")
|
||||
```
|
||||
And one SYS line (after "THE LAWS THAT MATTER MOST", `api_agent.py:290-314`):
|
||||
`3. If your card names a lever, grep the cookbook for the § its crib cites BEFORE drafting; a
|
||||
plateaued diff whose class says [permuter] is DONE — submit NEAR with the class, do not spend
|
||||
ten more compiles on a 2-ins register swap.`
|
||||
|
||||
### 6.3 `tools/build_wave_atlas.py:360` — attach the tell counts
|
||||
Join each card's member row to `.run/feat.<binary>.jsonl` (keyed by addr; already produced by
|
||||
`tools/atlas_features.py`) and attach
|
||||
`'tells': {'extpair': N, 'dupselect': N, 'magic_div': N, 'sign_lh': B, 'sign_lb': B}`;
|
||||
`_fuel` renders "expect N sll/sra pair sites / N repeated selects". Cost: one dict load per
|
||||
binary at build time.
|
||||
|
||||
### 6.4 `tools/ox_campaign.py:264` — per-lane budgets
|
||||
The tells lane needs `--max-turns 40 --max-cost-per-fn 0.40` (cards are 2.4× bigger and stack
|
||||
3–5 idioms; 36% of failures ended at the caps). Make both per-lane fields in the lane spec
|
||||
(`tools/lanes/campaign.sh:33` syntax) rather than constants.
|
||||
|
||||
### 6.5 Post-wave recovery (wave-closing step 2, per the standing 7-step sequence)
|
||||
- `bucket==permuter && closeness<=8` → `tools/permuter/` pass (zero tokens; 16 candidates now).
|
||||
- shape-MATCH + reloc-AGREE strays → §173 ladder; the 9 `main` ones go to the next gate_main
|
||||
slate instead of dying in the wave path.
|
||||
- `.run/s59_tells/open_drafts.json` + `open_nodraft.json` are the current worklists.
|
||||
|
||||
### 6.6 80+ band — stop drafting it wide
|
||||
Route extend-tell ≥ 80 ins through `tools/idiom_serial.py --lever extend-tell` (built for wall
|
||||
classes, never yet run on tells — its ledger holds only jtbl rows). Its distill step is the
|
||||
compounding channel the 2,000-way fan-out lacks, and `.run/idioms_learned.md` (currently absent)
|
||||
is where 4.1–4.4 should also land so the serial lane reads them.
|
||||
|
||||
## 7. The 75% verdict, per band (cumulative bank of carded tell-lever fns, live tree)
|
||||
|
||||
| band | banked now | verdict |
|
||||
|---|---|---|
|
||||
| 5–50 | 75/100 = **75%** | already there; retries converge without any change |
|
||||
| 50–80 | 83/144 = **58%** | **75% reachable** with §6 alone. 75% of 144 = 108 ⇒ 25 more banks needed from the 61 open (41% of failures). The measured ≤80 failure mix (26% shape-MATCH strays × 64% §173 recovery ≈ 17pp, 12% permuter-bucket × ~70% ≈ 8pp, ~22% one-construct LD × crib+budget ≈ 9–13pp, plus mid-gap partials) recovers ~35–45% of failures — 41% sits inside that band, tight but evidenced. No model upgrade required. |
|
||||
| 80–120 | 34/118 = 29% | 45–55% plausible (crib + budgets + permuter + recovery); **75% not evidenced** — the LD gap-mass shifts to multi-region misreads (54/140 > 8 words) |
|
||||
| 120+ | 26/93 = 28% | wide drafting is the wrong tool; serial lane + remap or park. 75% out of reach at current drafter tier |
|
||||
|
||||
Instruction-weighted honesty: the lane's 86,602 ins sit mostly ABOVE 80 ins (atlas means: extend
|
||||
76,972/1,010 ≈ 76, but the open residue skews large — the 390 still-open average 103). So "the
|
||||
lane at 75%" is true only as a ≤80-ins statement; the *instruction* frontier of this population
|
||||
is a serial-learning problem, not a wave-tuning problem. The three cracked exemplars each carried
|
||||
3–5 stacked laws; that stacking is the honest reason tells collapse above 80: p(all-k-idioms)
|
||||
falls geometrically, and only a compounding lane (or the §6.1 cribs raising per-idiom hit rate)
|
||||
bends it.
|
||||
|
||||
**Refuted along the way** (do not re-spend): §235 symbol-identity as the tells failure mode
|
||||
(body-wrong confirmed: only 6/66–5/47 shape-matches among MISMATCH?); "drafts miss the extend
|
||||
pairs" (74% have exact pair counts); "the models don't use the cookbook" (650/796 grepped — with
|
||||
the wrong words).
|
||||
@@ -310,11 +310,50 @@ THE LAWS THAT MATTER MOST:
|
||||
A draft that calls the WRONG function reports a clean MATCH. After MATCH, walk the .s once more
|
||||
and check every symbol you wrote against its own relocation lines.
|
||||
2. A seed or twin gives you the SHAPE, never the symbols, and never the literals. Re-read both.
|
||||
3. IF YOUR CARD NAMES A LEVER, grep the cookbook for the § its crib cites BEFORE you draft. And a
|
||||
plateaued diff whose class says [permuter] is DONE: submit NEAR with the class rather than
|
||||
spending ten more compiles on a 2-instruction register swap the search cannot reach from C.
|
||||
4. AN ARGUMENT YOU INVENTED CHANGES SCHEDULING (§263). If your only residual is one stolen delay
|
||||
slot holding a `move $aN,$sN`, you gave a no-argument callee a parameter — check its arity
|
||||
against decl_prior before touching a fence or the permuter, which cannot change call arity.
|
||||
|
||||
Work until match_one says MATCH, or until you have genuinely exhausted your levers -- then submit
|
||||
with an honest status and say what the cookbook did NOT already tell you."""
|
||||
|
||||
|
||||
# THE LEVER CRIB (P31 S59). The card used to state its lever as a bare label and nothing else, and
|
||||
# the label is a word this project's 750-section cookbook DOES NOT CONTAIN: measured over the tells
|
||||
# waves, 108 failure transcripts grepped `extend-tell` and 28 grepped `swaprepeat` for zero hits
|
||||
# while the knowledge sat at §172a/§172b/§264 under different words. A label the drafter cannot look
|
||||
# up is not fuel; it is a dead end that costs a grep and buys nothing. Each entry below states what
|
||||
# the tell MEANS, names the § to grep, and gives the C spellings — all byte-proven (§264).
|
||||
LEVER_CRIB = {
|
||||
'extend-tell': (
|
||||
"extend-tell = §172b-1 + §264 (grep 'sll/sra-16 pair tell'). The target promotes\n"
|
||||
"REGISTER-held s16 values. Four placements, four C spellings:\n"
|
||||
" * naked sll/sra pair = a MULTI-DEF s16/u16 variable merged across branches — write it\n"
|
||||
" on BOTH paths mirroring the mem slot (`x[i] = w = ...`), then compare (s16)w;\n"
|
||||
" * plain lh = one promoted use — read the s16 memory DIRECTLY, no temp;\n"
|
||||
" * sll before a jal, sra IN its delay slot = the (s16) cast written INLINE in the call\n"
|
||||
" argument (a temp kills the delay-slot fill);\n"
|
||||
" * lhu + separate sll/sra = the value is ALSO used as u16 — keep the u16 variable.\n"
|
||||
"Sign test (s16)t < 0 emits sll+bgez only (the sra is elided).\n"
|
||||
"NO register pins first — §17 is the last resort; pins block the parm-copy and arg-copy\n"
|
||||
"shapes. If a parameter is copied to $aN, type the PARAMETER and use it everywhere rather\n"
|
||||
"than inventing `self = arg0`.\n"),
|
||||
'swaprepeat-tell': (
|
||||
"swaprepeat-tell = §172a + §172b-2 (grep 'swapped-arm select tell'). Textual REPEATS are\n"
|
||||
"load-bearing: a repeated compare/select re-emits its own slt+branch, so write the repeat\n"
|
||||
"TEXTUALLY — nested macros repeat their operands, inline functions collapse them.\n"
|
||||
"A range check is plain `if (-lim < x && x < lim)`; the second slt rides the first beqz's\n"
|
||||
"delay slot for free.\n"),
|
||||
's16-div-tell': (
|
||||
"s16-div-tell = §172b-4 (grep 'Division-by-constant sign-correction'). For s16/K magic\n"
|
||||
"multiplies keep the promotion as an EXPRESSION, never a temp, so the sign-sra can reuse\n"
|
||||
"the <<16 intermediate; the magic pairs for K are in §167-25 and the odd-part table.\n"),
|
||||
}
|
||||
|
||||
|
||||
def _fuel(t, card):
|
||||
"""THE CARD FUEL, in the wave's own proven wording (P31 S56).
|
||||
|
||||
@@ -383,6 +422,14 @@ def _fuel(t, card):
|
||||
out.append( "WORTH, not where an answer is.")
|
||||
if card.get('lever'):
|
||||
out.append(f"\nThe atlas labels this card's lever: {card['lever']}.")
|
||||
crib = LEVER_CRIB.get(card['lever'])
|
||||
if crib:
|
||||
out.append(crib)
|
||||
out.append("Universal fold-blockers when a shape refuses: name the inner term of a "
|
||||
"nested +/- as a temp (keeps association); hold a non-slti constant bound "
|
||||
"in a LOCAL assigned right before its `if` (defeats the 0xFFFF "
|
||||
"canonicalization and lets the arm's store reuse its register); hold a "
|
||||
"global base in a pointer local when it must live in an $sN across calls.")
|
||||
return '\n'.join(out)
|
||||
|
||||
|
||||
|
||||
+23
-5
@@ -187,7 +187,17 @@ def shard_targets(tag, cards_path, workers):
|
||||
return targets
|
||||
|
||||
|
||||
def draft(tag, lanes, maxtok, max_turns, ramp_seconds=float(os.environ.get('RAMP_SECONDS', '240'))):
|
||||
# PER-LANE AGENT BUDGETS (P31 S59, measured). The turn/cost caps were one global pair, and the
|
||||
# tells lane's cards are 2.4x the default lane's size (median 89-95 instructions vs 37-39) and stack
|
||||
# 3-5 idioms each: 98 of 270 final tells attempts ended AT the turn cap and 36% of failures ended at
|
||||
# a cap of some kind — i.e. they ran out of budget mid-work rather than failing. A budget that fits
|
||||
# a 38-instruction card starves a 95-instruction one. {lane: (max_turns, max_cost_per_fn)}.
|
||||
LANE_BUDGET = {'tells': (40, 0.40)}
|
||||
LANE_BUDGET_DEFAULT = (None, 0.15) # None = the campaign's --max-turns
|
||||
|
||||
|
||||
def draft(tag, lanes, maxtok, max_turns, ramp_seconds=float(os.environ.get('RAMP_SECONDS', '240')),
|
||||
wave_lane=None):
|
||||
"""Draft one wave across SEVERAL MODEL LANES at once.
|
||||
|
||||
Concurrency on ONE model is spent (P31 S58, measured): every 429 we have ever recorded is
|
||||
@@ -200,6 +210,11 @@ def draft(tag, lanes, maxtok, max_turns, ramp_seconds=float(os.environ.get('RAMP
|
||||
owns a known slice, and every request already carries its model in the telemetry — so the
|
||||
per-lane 429 rate is measurable after the fact instead of assumed.
|
||||
"""
|
||||
_t, _cost = LANE_BUDGET.get(wave_lane or '', LANE_BUDGET_DEFAULT)
|
||||
_turns = _t if _t is not None else max_turns
|
||||
if _t is not None:
|
||||
log(f" lane {wave_lane}: budget {_turns} turns / ${_cost:.2f} per function "
|
||||
f"(default {max_turns}/$0.15)")
|
||||
outdir = f".run/wave_{tag}"
|
||||
shutil.rmtree(outdir, ignore_errors=True)
|
||||
os.makedirs(outdir, exist_ok=True)
|
||||
@@ -261,7 +276,7 @@ def draft(tag, lanes, maxtok, max_turns, ramp_seconds=float(os.environ.get('RAMP
|
||||
procs.append(subprocess.Popen(
|
||||
[PY, "-u", "tools/api_agent.py", "--targets", tf, "--cards",
|
||||
f".run/wave_{tag}_cards.json", "--out", f"{outdir}/shard{idx}",
|
||||
"--max-turns", str(max_turns), "--max-cost", "1.0", "--max-cost-per-fn", "0.15"],
|
||||
"--max-turns", str(_turns), "--max-cost", "1.0", "--max-cost-per-fn", str(_cost)],
|
||||
stdout=fh, stderr=subprocess.STDOUT, env=env))
|
||||
started[model] += 1
|
||||
# STAGGER THE LAUNCH. Every shard issues its first request immediately, so starting N at
|
||||
@@ -615,7 +630,8 @@ def run_drafter(a):
|
||||
log(f" {tag}: draw failed — retrying next cycle"); time.sleep(30); continue
|
||||
targets = shard_targets(tag, cards, a.workers)
|
||||
t0 = time.time()
|
||||
procs = draft(tag, parse_lanes(a.models, a.model, a.workers), a.maxtok, a.max_turns)
|
||||
procs = draft(tag, parse_lanes(a.models, a.model, a.workers), a.maxtok, a.max_turns,
|
||||
wave_lane=lane['name'])
|
||||
|
||||
# PRE-DRAW THE NEXT WAVE **AFTER** THE SHARDS ARE RUNNING, never before. Drawing takes
|
||||
# minutes of CPU (build_wave_atlas over the whole atlas), and doing it between the draw and
|
||||
@@ -806,7 +822,8 @@ def main():
|
||||
log(" draw failed — skipping wave"); continue
|
||||
targets = shard_targets(tag, cards, workers)
|
||||
log(f" {len(targets)} targets sharded")
|
||||
procs = draft(tag, parse_lanes(a.models, a.model, workers), a.maxtok, a.max_turns)
|
||||
procs = draft(tag, parse_lanes(a.models, a.model, workers), a.maxtok, a.max_turns,
|
||||
wave_lane=lane['name'])
|
||||
inflight = (tag, cards, targets, procs, time.time())
|
||||
|
||||
tag, cards, targets, procs, t0 = inflight
|
||||
@@ -824,7 +841,8 @@ def main():
|
||||
ncards = draw_wave(ntag, a.cards_per_wave, nband, nlane.get("levers"))
|
||||
if ncards:
|
||||
ntargets = shard_targets(ntag, ncards, workers)
|
||||
nprocs = draft(ntag, parse_lanes(a.models, a.model, workers), a.maxtok, a.max_turns)
|
||||
nprocs = draft(ntag, parse_lanes(a.models, a.model, workers), a.maxtok, a.max_turns,
|
||||
wave_lane=nlane['name'])
|
||||
inflight = (ntag, ncards, ntargets, nprocs, time.time())
|
||||
else:
|
||||
log(f" wave {ntag}: draw failed — next cycle will redraw")
|
||||
|
||||
Reference in New Issue
Block a user