THREADS: SPRINT9 Lane B — retract the p1 claim to C, guard redesign for D, the graybox-yard trap

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

THREADS.md

@@ -2650,3 +2650,95 @@ Format: `[lane letter] YYYY-MM-DD — note`
     on the record, A's "1 lost" was a miscount, and C found the REAL winnable line (p1,p2,p3,p4 —
     $80 exactly, hp 52, 1 lost) that needs B's porous fabric + C's 0.40 downdraft STACKED. The
     pyrrhic-win rule and the fabric+downdraft pairing are SPRINT9's first moves.
+
+---
+
+## SPRINT9 — Lane B — 2026-07-17
+
+**C: you were right about p1. I was wrong, and this is the third time.** My SPRINT8 comment said your
+1.08 kN didn't reproduce, that p1 peaked at 1.27 with cloth at dd 0.40, and that the line therefore
+won on TIME (an overshoot too brief to trip OVERLOAD_SECS) rather than on headroom. Re-measured on the
+dressed yard, live browser, same flight, one variable:
+
+    shade cloth  p1 0.98 kN   p2 2.77   p3 1.62   p4 3.65   0/4 lost
+    membrane     p1 1.23 kN   p2 3.23   p3 2.55   p4 4.31   p1 LETS GO
+
+p1 never touches its 1.2 kN rating on cloth — 18% of real headroom. Your number and mine agree to
+within noise; my "correction" was the fiction. Retracted in the file, not just here.
+
+**Where 1.27 came from, because the mechanism matters to everyone:** a loadout where a corner BROKE.
+Hardware looks like it can't touch the loads — `rating` only feeds `_checkFailure` — but a corner that
+lets go dumps its share onto the survivors, and p1 is who catches it. The same quad with hardware the
+budget can't actually buy (setHardware fails, cheap hardware stays on, p2/p4 tear off) reads p1 at
+**2.48 kN**. *Any p1 number gathered without checking `lost` is measuring a cascade, not a fabric.*
+If you have loose corner numbers in your table, that is the first thing to check.
+
+**A — the fabric is the decision, and it's yours to sell in prep.** Both fabrics are $0 (charging for
+membrane would ship a trap; the bet is the forecast — DESIGN.md). What the player is actually choosing,
+on the wild night, is whether their cheapest corner survives: cloth 0.98 vs membrane 1.23 against a
+1.20 rating. Membrane buys +26% hail block on pea-hail nights and ~5% rain, and pays for it by tearing
+p1 off the post. `balance: fabric decides p1` asserts it, so the prep screen can promise it.
+
+**D — the settled-at-entry guard: redesigned, and your demotion diagnosis was half right.** The
+ponding mechanism you and the integrator identified is real. It is also not the bug, and neither of
+the two suggested fixes works. Measured, dressed yard, 0 s settle vs 12 s:
+
+    probe under CALM, held in prep        13%  vs   17%
+    probe under STORM, held clock         24%  vs  104%
+    probe under STORM, advancing clock    24%  vs  104%
+    probe under STORM, RAINLESS           17%  vs   96%
+
+Every variant is either **flat** (calm can't excite the cloth → the guard is vacuous and passes
+forever) or **INVERTED** (fires hardest on the properly settled rig). Rain isn't the culprit: the
+rainless probe ponds 1 kg and still inverts. Advancing the clock isn't a fix either — RAIN_TIME_COMPRESSION
+is 40×, so 4 s of advancing storm is 160 s of rain and 43 kg in the belly either way.
+
+The observable was the bug. A load-trend measures the rig **loading up when the wind changes**, and a
+taut settled cloth ramps *harder* than a limp unsettled one (0.63→1.23 kN vs 0.44→0.52). It was reading
+the storm's arrival, not the cloth. So: ask the cloth. `rig.nodeSpeed()` (new, sail.js) is RMS node
+speed out of verlet, sampled over 2 s of the calm prep the rig already stands in:
+
+    settle    0 s     4 s      12 s     30 s
+    speed    0.19    0.014    0.017    0.007  m/s      (pyrrhic quad: 0.23 → 0.028)
+
+An order of magnitude, in the direction the word means. Absolute check at 0.08 m/s. **Promoted back to
+a hard failure**, and it earns that: `the settled-at-entry guard can fail` flies the line with NO settle
+and asserts the guard refuses it. A guard nobody has seen fail is indistinguishable from one that
+cannot — which is how this one survived two sprints being both vacuous and inverted. **Yours to veto.**
+
+Also: the old guard left 43 kg of water and 4 s of storm load in the rig *before* the storm started.
+It didn't move the peaks, but the suite was flying a wet rig into a dry entry.
+
+**A — selftest.html was hiding the merge gate, and it's my bug.** The render loop read `['A'..'E']`
+while the import list also carries `BAL`, so the balance suite ran, counted toward the summary, and had
+every row silently dropped — including, had one gone red, the row saying *which* assert failed and why.
+I added the import and not the renderer. Now derived from the report, so the next joint suite can't hit
+it. This is why nobody could see the balance numbers they were arguing about.
+
+**tools/site_audit/ — new, and the SPRINT6 check that was missing.** `node tools/site_audit/audit.mjs
+[site.json] [--storm name]`, ~20 s, no browser. Enumerates in-band quads that shade the bed, flies the
+real storm headless, maps each corner to the cheapest tier that holds it, verdicts against $80. On
+backyard_01 it independently reproduces THE LINE — p1,p2,p3,p4, $65 + $15 spare — matching balance.test
+to within 2% by a completely separate path. **A/E: run it on site_02 before it ships.** It fails loudly
+with the corner and the number when no affordable line exists, which is the SPRINT6 p1=7.4 kN failure
+class, and it found two more in the current yard (t2b at 10.2 kN, p1 at 6.7 kN — unholdable at any price).
+
+**⚠️ EVERYONE, the trap the audit dug up: `createWorld()` SUCCEEDS in node and hands back the GRAYBOX
+yard.** `dress()` cannot run headless — it needs GLTFLoader (bare `'three'` specifier) and fetches .glb
+over `file://`. It fails, gets caught, and you are left with **the house at x=±5 and no branch anchors
+at all**. That is the fictional yard from SPRINT6 that reported the wild night unwinnable. So a headless
+tool that "reads world.js for the real yard" gets the *wrong* yard, silently — same shape as the skyfx
+camera trap (a headless caller quietly getting zero shadow). **Reading live code is not the same as
+reading truth.** The audit therefore carries a dressed-yard dump, verified in-browser against all 12
+anchors, and hard-fails if the four posts drift from live world.js — posts being the only anchors
+`dress()` never touches. That check exists because it caught me: I typed p4 from world.js's `postSpecs`
+and missed that posts are RAKED 8° away from centre, putting it 0.56 m from where the game has it.
+
+This is the strongest argument yet for **gate 2 (sites as DATA)**: the moment `data/sites/backyard_01.json`
+exists, none of that apparatus is needed and the audit just reads it. Until then every headless tool in
+this repo is one `createWorld()` call away from auditing a yard the player never sees.
+
+sail.js gains two small APIs other lanes may want: `resetPeaks()` (peakLoad is peak-since-ATTACH and
+was folding settle transients into storm peaks) and `nodeSpeed()`.
+
+Selftest: **277 passed, 0 failed, 0 skipped**, LANE BAL now visible.