The operator's play report on FieldShieldTest ("weak hits are very
janky, clicking outside the hex still makes it tremble and shatter,
pooling seems very rough") traced to two provable defects, not one
tuning constant:
- ShieldPanel.Hit had no bounds check at all. A point handed to it was
always enqueued and injected regardless of how far outside the
panel's own panelRadius it was.
- ShieldInterceptor.FindReusable decided reuse from a raw world-space
chord distance to the panel's stale spawn-time Position, using
PanelRadius*ReuseFraction (authored up to 2x) as the threshold. On a
non-uniformly scaled/curved envelope a chord across curvature is not
the panel's own in-plane extent, so a click clearly outside the
rendered tiling could still measure inside threshold and get routed
to Hit() -- which had nothing to reject it.
- Reusing a panel never refreshed its pooled DieAt, so a panel struck
again could still expire and get returned to the pool mid-fracture
("pooling seems very rough").
Fix: ShieldPanel.ContainsWorldPoint is now the one owner of "is this
point within my own radius" (panel-local in-plane distance vs.
panelRadius), used both as Hit's hard gate and as FindReusable's
candidacy test (with ReuseFraction, now capped at 1, narrowing but
never widening past the physical radius). Reuse now also extends
DieAt from the moment of the new hit.
Pinned in ShieldPanelCut3Verify.CheckHitOutsideRadiusDoesNotReachPanel:
a direct Hit() and a routed Absorb() outside a panel's own radius must
enqueue nothing and leave its state buffer untouched, and a genuinely
outside click opens its own panel. Deterministic (checks the pending
queue and no-op state, not temper drift) because breakBudgetFraction
paces fracture over many frames on purpose, so temper keeps drifting
from an earlier hit alone -- a drift-based check would have been
measuring that, not this rule (caught this against the check's own
first run).
All four ShieldPanelCut<N>Verify probes pass (batchmode, real GPU
device, no -nographics).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Cut 5's own probe passed while breaking Cut 3's reuse rule: with R7 holding
the rim, the through-thickness echo (KEcho/_EchoReflect) is the only
remaining source of sub-critical tension for a weak hit, and its magnitude
(-0.65) was weaker than the still-live residual compression velocity left
by the direct KInject kick by the time the echo fired. The echo's own
header comment says a free-surface reflection inverts with unit
magnitude, matching the pre-R7 rim's Dirichlet condition -- -0.65 was a
lossy fraction that could reduce but never flip the sign of u, so a weak
hit's u never crossed positive and KFracture's `u <= 0` guard skipped the
sub-critical erosion branch forever, no matter how many hits landed.
Raised echoReflect from -0.65 to -0.9 (ShieldPanel.cs) -- enough margin to
flip a weak hit's residual compression into net tension without needing a
delay/seeding change; echoDelaySteps and KInject's seeding were not the
defect. All four probes now pass, including Cut 3's reuse rule (temper
0.4180 -> 0.3997 on a reused panel).
Also lands the multi-hit-reuse rule Cut 5's own report flagged as unpinned:
CheckRepeatedWeakHitsBreakWhereOneDoesNot in ShieldPanelCut5Verify.cs.
Energy 15 (not 8, whose peak tension of ~0.14 sits under tensileStrength
0.2 and can never break at any repetition) breaks on hit 2 of a live,
reused panel while a cold single hit -- given equal idle time (720 steps,
matching the multi-hit run's total budget, to rule out "just needed more
time to ring out") -- never breaks. This is the Cut 3 reuse rule and the
temper mechanism producing the operator-facing payoff together.
Verified: ShieldPanelCut1Verify, Cut2Verify, Cut3Verify, Cut5Verify all
PASS via batchmode (graphics device required, not -nographics).
Operator watch: echoReflect at -0.9 is close to the physical "unit
inversion" ceiling implied by the header comment; if later tuning raises
hit energies or lowers damping, watch for the echo overshooting into an
immediate break on a single moderate hit (observed at -1.4 and even -1.0
in earlier tuning passes) rather than the intended multi-hit erosion.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
R7 (operator ruling): the panel's emitters hold its edge, so the rim is no
longer a free surface. Deletes the openW ghost term end to end -- its only
consumer in ShieldSim.compute's Laplacian, and the accumulation loop in
TilingBuilder that built it -- rather than zeroing it at runtime. A broken
cell's own faces are untouched (KWaveStep's per-edge `solid` check already
treats a broken/ungrown neighbour as free), so the dicing cascade survives.
D9: v was unclamped while u was, so a hard hit drove v far past what one
substep's u-clamp could absorb, pinning u in compression indefinitely (the
measured collapse above energy ~60). Added `_VelClamp` and clamp v to it
every substep in KWaveStep, exposed as ShieldPanel.velocityClamp (100).
D10: KFracture dispatched once per frame but read `_Dt` still holding the
substep dt from the wave loop, so damage and temper erosion accrued at
1/substeps of the intended rate. Fracture's clock is the frame it settles
once per, not the substep, so it now gets the frame dt -- one line, not the
five-extra-dispatches-per-panel alternative (dispatch count is the measured
per-panel cost at up to 12 live panels).
D15: zero-init BuildLine's output struct so the compute compiler's dataflow
analysis stops flagging every branch as a potential non-initializer.
Tuning (ShieldPanel.cs Material header): tensileStrength 0.9->0.2,
temperErosion 2.2->12, damageGain 6->100, breakAt 0.75->0.08,
storedEnergyGain 0.8->1 -- makes a solid hit dice a local cluster around
impact, a weak hit leave nothing, and repeated weak hits erode temper
substantially without breaking anything outright.
Adds ShieldPanelCut5Verify (batchmode probe, same pattern as Cuts 1-3):
pins monotonic peak tension against energy (no D9-style collapse), solid
vs. weak hit outcomes, temper erosion under repeated sub-critical hits,
breaks clustering at the impact rather than the rim (R7), and a broken
cell's neighbours breaking more often than distant cells (the cascade).
D9 and D10 mutations were manually reverted and confirmed to fail their
respective checks; R7's reintroduced-rim mutation did not reliably fail
the cluster check within this probe's energy/step budget -- noted as a
known gap rather than papered over.
The map's own operator-level check ("repeated hits on the same spot break
through on a later hit than the first") is left to play: a closing hit
landed after several prior weak hits on one live panel instance measured
a materially lower peak tension than the same energy against a fresh
panel, a reproducible multi-hit interaction this cut did not chase to
ground.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
D17: KInject/KWaveStep/KFracture gate on growth, which only KUpdate writes and
only after this same frame's injection dispatch, so a panel struck the instant
it spawns swallowed its own creating hit. ResetSim now seeds bState directly
from a precomputed per-cell arm state (backdated by one growDuration) instead
of leaving KInit's zero in place, so the strike centre reads as already grown.
D20: _PanelSpawnTime was hardcoded to 0f every frame; it now tracks each
panel's own backdated arm instant, so growth keeps advancing correctly on the
frames after spawn instead of collapsing back under the gate.
Pool-rebuild trap (not map-numbered, but load-bearing for the whole pool
argument in §2.4): OnEnable unconditionally rebuilt and OnDisable
unconditionally released, so every SetActive cycle a pooled panel goes through
paid the 1.3-21ms Rebuild cost the pool exists to avoid. OnEnable now only
rebuilds when not yet built; release moves to OnDestroy. ResetSim is also
self-healing (builds if not yet built) rather than depending on OnEnable's
timing, which a batchmode probe showed is not synchronous with
Object.Instantiate in every context this runs in.
D21: Penrose's iteration count was CeilToInt(...)+1, always overshooting the
requested cellSize by a full extra phi-subdivision. Rounding to the nearest
achievable iteration count instead brings Penrose within 2-5x of Hex's cell
count at the same radius/cellSize, down from an order of magnitude.
TilingBuilder.cs gains a CellState mirror struct (matching ShieldCommon.hlsl's
CellState field-for-field) so ResetSim can write growth/temper/breakTime
directly and the Cut 3 verification probe can read them back without a new
readback path.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Cut 1 of docs/shield-panel-cut.md: land ShieldPanel.cs, ShieldPanel.shader,
ShieldSim.compute, ShieldCommon.hlsl, TilingBuilder.cs and Tilings.cs under
Assets/Shaders/Compute/ShieldField/, alongside Lightning/ and Slime/.
Port (URP -> built-in, three substitutions per the map's §2.1 probe):
- drop the UniversalPipeline SubShader tag, add IgnoreProjector
- swap the URP Core.hlsl include for UnityCG.cginc
- TransformWorldToHClip -> UnityWorldToClipPos (3 call sites)
Drawing collapse (D2/D5): RenderParams has no shaderPass member in
6000.3.24f1, and a multi-pass material draws every SubShader pass from one
indirect call regardless, so the three per-pass args buffers and draw calls
collapse to one args buffer sized to the widest pass (Outline) and one
Graphics.RenderPrimitivesIndirect call. Deleted: argsOutline, argsConduit,
drawFill/drawOutline/drawConduits and their branches.
Also deleted: ShieldPanelTester (the click harness the operator ruled
against; it also can't run — activeInputHandler is New Input System only)
and [ExecuteAlways] (incompatible with a pool per D13; Rebuild stays on the
context menu).
Defects fixed:
- D1: TilingBuilder.cs's inner `dist` shadowed the outer one (CS0136);
renamed to nbrDist.
- D3: CellStatic was 15 floats against a declared Stride of 16; added pad2
on both the HLSL and C# sides.
- D4: `centroid` is reserved in the compute compiler; renamed the
ShieldCommon.hlsl field to `center` (kept the C# field name as-is; it's
not reserved there and the two sides don't need matching names, only
matching layout).
- D6: fixed the comment on _BreakBudget — it's two elements (count + cap),
not the "single element" the old comment claimed.
- D12: the two-element uint[] Update() allocated every frame is now a
cached field, cheaper still under a pool of many live panels.
- D18 (R5 — operator ruling, not deferred to a follow-up): Penrose ignored
panelRadius entirely — a fixed authored iteration count meant the tiling
was just uniformly rescaled, not resampled, so cell count never grew with
the panel. Every other tiling here uses cellSize to hold local cell size
constant while radius extends coverage; Penrose now derives its iteration
count the same way (each substitution divides edge length by phi, solved
for the count that brings a unit edge to ~cellSize once scaled to
radius). This retires the now-meaningless `penroseIterations` field and
the `subdivisions` parameter it threaded through Tilings.Generate.
- Also removed a dead `idx` counter in Tilings.Voronoi (D8's mention of
dead code; the Vector2.zero sentinel ambiguity itself is unfixed, same as
D7/D19 — Voronoi and TriHex stay recorded defects, not fixed, per R5).
Not fixed here, per the map: D7 (Voronoi is unusably slow) and D19 (TriHex
has zero dual edges) are recorded, not fixed. D9/D10 (fracture tuning) are
Cut 5. D17 (growth gate blocks the spawning hit) and D20 (_PanelSpawnTime
hardcoded to 0) are Cut 3.
Repo-specific fix not in the map: Assets/Scripts/UI/ContextMenu.cs defines
a global `ContextMenu` MonoBehaviour that shadows UnityEngine.ContextMenu,
so `[ContextMenu("Rebuild")]` doesn't compile here (CS0616) even though it
did in the map's throwaway probe project. Used the fully-qualified
`[UnityEngine.ContextMenu(...)]` instead.
Adds Assets/Scripts/Editor/ShieldPanelCut1Verify.cs, a batchmode probe (no
NUnit path exists per the map's Q7 — Tests.asmdef can't see
Assembly-CSharp): asserts zero shader/compute compile errors and that
Penrose cell count actually scales with panelRadius (D18). Also adds
tests/mutation_tests_shield_panel_cut1.py, which mutates the D18 fix, the
D4 fix, and the shaderPass deletion in turn and asserts the probe goes red
each time, restoring the source after each case (own file per the
convention mutation_tests_addressables_cut.py established, so as not to
touch item-provenance's or the addressables cut's harnesses).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Meta Vens