[screen] round 2 notes + snapshot-access contract request

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

fktry/lanes/LANE-SCREEN.md

@@ -149,3 +149,79 @@ express — moiré needs a real interference lattice, ringing needs edge-detecte
 macroblock bricks want the 8×8 grid to be *legible*. That's new shader passes, not new
 weights. Also THE SCREEN should probably show the shipped item's silhouette for a beat as it
 lands. Say the word on round 2.
+
+### Round 2 — 2026-07-17 — Opus 4.8
+
+SHIPPED: orders 1–3 done and verified; order 4 (optional) deliberately deferred with the
+contract request filed below.
+- `params.ts` — PARAMS extended 8 → 11 (`block`, `ring`, `moire`). The original eight keep
+  their prototype tuning untouched. Because the ledger/easing/wobble/stats/tests all iterate
+  PARAMS, the three new classes inherited the whole machine for free.
+- `glitchStack.ts` — three new passes + the two overlays.
+- `overlays.ts` (new) — stamp placard + scram bar painters.
+- `index.ts` — stamp queue/lifecycle, scram trigger, debug hook now behind `import.meta.env.DEV`.
+- `corruptionMap.test.ts` — 30 tests green (was 24).
+
+DECISIONS:
+- **The block lattice is locked to screen space (`vUv`), not to the warped `uv`.** That
+  rigidity is the whole point: the grid stays a hard lattice while melt/mosh drag the picture
+  underneath it. Quantized in a 80×45 grid = true 8×8 source pixels at 640×360.
+- **Ringing needed a second attempt.** v1 sampled the gradient at ±1.6px, so halos clung to
+  the 1px boundary and the whole thing read as "sterile, but sharper" — it flunked your
+  legibility bar and I nearly shipped it. Fixed by sampling the gradient WIDE (out to 7px) so
+  the halo has room to ring out, and running the oscillation *parallel* to the edge that spawned
+  it. Mean abs pixel diff vs sterile went 16.5 → 25.5, and it now reads as haloing at a glance.
+- Overlays composite in screen space so a melting feed never drags the placard — a stamp that
+  melts stops being a stamp.
+- Stamp queue caps at 2 pending. One stamp per EVENT, 0.4s each, size scaled by count
+  (0.72→1.0 across 1..40 units) with an overshoot on impact. A burst reads as a couple of
+  stamps, never a strobe.
+- Scram bar position is derived from `entity`+`tick`, not `Math.random()`, so a given run
+  always wipes the same way (I'm not bound by the determinism rule, but free determinism is
+  worth having when debugging a report).
+- `coefficient-pack` moved onto `block` (per your "coeff overdrive must READ as a blocky
+  grid"); `gibbs-wraith` moved onto `ring`. New tests pin bricks-vs-slurry and assert the three
+  new classes carry zero of each other's signature.
+
+VERIFIED (drove the real module on the real canvas, read GL pixels, screenshotted contact sheets):
+- **Bricks**: within-cell variance on the 8px lattice drops 689 → 236 (cells genuinely
+  flattening = quantization, not noise). Reads as an unmistakable quilted posterized grid.
+- **Ringing**: bright oscillating ghost rings hugging every bar edge at 3.2× zoom; flat sterile
+  edges for comparison. MAD vs sterile 25.5.
+- **Moiré**: diagonal rainbow interference lattice over the whole feed. MAD vs sterile 35.3.
+- **Distinctness** (mean abs pixel diff): bricks↔slurry 21.9, ring↔moiré 29.5, moiré↔slurry 34.3.
+  Nothing looks like anything else.
+- **Stamps**: MELT ×40 punches a big orange placard, fades by 0.33s, gone by 0.6s; HF DUST ×1
+  stamps visibly smaller in its own chip colour; stamp stays rigid while the feed melts under it.
+- **Scram**: torn white TRACKING LOST bar on exactly one frame, absent the next.
+- **Budget: 0.047 ms/frame wall-clock with a forced GPU sync** (`readPixels` flush), measured
+  over 120 frames — and *identical* for sterile vs. every artifact class lit at once plus
+  brownout. The new passes are free at this resolution. CPU-only rolling avg: 0.003–0.02 ms.
+
+BLOCKED/BROKEN: nothing. Two notes for the record:
+- The verification tab is `hidden`, so rAF never fires and the main loop sits frozen at ~TICK 8.
+  That is the harness, not the app — I drive `frame()` with synthetic timestamps instead, which
+  is also why my numbers are reproducible. Related: a hidden tab clears the WebGL drawing buffer
+  after compositing, so a plain screenshot races the clear and sometimes catches pure black. It
+  cost me a while to prove that was the harness and not a moiré bug. Contact sheets (drawImage
+  into a 2D canvas in the same JS task) are immune; I'd recommend the trick to any lane doing
+  canvas verification.
+- Mid-round I saw reds in `src/sim/sim.test.ts` and `src/render/index.ts` tsc; both cleared as
+  the lanes landed. HEAD is clean repo-wide, 30/30 mine.
+
+CONTRACT REQUEST: **`frame(timeMs: number, snap?: SimSnapshot)`** — taking you up on the v3
+offer. Events give me edges (`brownout` on/off); they can't give me a *continuous* quantity.
+Signal strain wants THE SCREEN to tremble as `bandwidth.stored` drains toward zero — dread
+before the brownout, not just the seizure. Same seam would cover reacting to aggregate `heat`.
+Read-only, no mutation, and I'd keep the snapshot un-retained across frames per the invariant.
+If granted I'll build it round 3; the corruption ledger stays event-driven either way.
+
+PROPOSAL: the scram wipe at exactly one frame (as ordered) is ~16ms and honestly borderline
+subliminal — it reads more as a flinch than a legend. The awakening gets away with it because
+it's full-screen white; a bar doesn't. Suggest 3 frames (~100ms) so TRACKING LOST is actually
+readable. Trivial change, your call — shipped at one frame as ordered.
+
+NEXT (not self-assigning): stamps currently say the item's NAME; once MODELBEAST assets land I'd
+rather stamp the item's *silhouette* in its chip colour. Also the 3 new passes are all
+full-frame — moiré especially wants to grow from a *region* (the codex has it spreading as maze
+walls across the factory floor), which would want a spatial mask rather than a scalar.