M49: Crazy Fruits (match-3) — single-source engine, masked DQN (+57% over random), touch+mouse web game - #38
Merged
Merged
Conversation
4-agent investigation (game provenance, repo integration, conventions/input, match-3 RL prior art) -> docs/prd/CRAZY_FRUITS_PRD.md + PLAN.md M49. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…cade + tests - crazyfruits_solver.pg: 8x8/6-fruit match-3 rules (swap legality, cascades with proportional scoring 10*(k+1) + line bonuses, gravity/refill, deadlock reshuffle), Schrage minstd RNG (exact i32 both sides), 448-float observation (6 one-hot planes + would-match plane), scripted baselines (random/greedy/expectimax-1) and the PgCfDuelingNet forward - all shared C#/TS; pgconfig routes the TS twin to ClientApp/src/app/crazy-fruits. - CrazyFruitsBoard: public facade (state serialization, LoadGrid for directed tests). - Gate GREEN: 16 tests - init/post-move invariants (stable board, always a legal swap), mask == brute force, hand-computed 3/4/5-line + 2-step-cascade scores, reshuffle multiset preservation, minstd 10,000th-draw reference (399268537), determinism + state round-trip; C#<->TS parity: seeded 1,000-move random episode checksummed per-move over action/points/full grid -> identical (78377593, score 70990) against the generated TS twin under node type-stripping. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- CrazyFruitsEnv: IEnvironment + IActionMaskProvider + IStatefulEnvironment over the .pg core; reward = points/30, truncation (never termination) at the 30-move budget, full state round-trip, viz labels. - CrazyFruitsDqnCampaign on the DqnScoreCampaign spine (masked Double+Dueling, the Snake recipe re-pointed) + AddCrazyFruitsDqnCampaign() + `--game crazyfruits` (`--baselines N` prints the gate table; `--net` adds the trained net to it). - Gate GREEN: baseline ordering over 500 seeded episodes - random 2259.7+/-49.9, greedy 2387.0+/-49.3 (CI-separated), expectimax-1 4270.9+/-98.3 (cascade planning is the skill; line size is nearly irrelevant); campaign resume->train->checkpoint-> resume contract on the real env; one end-to-end 3k-step chunk trained and saved a deployable net. 21 CrazyFruits tests green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… watch scaffolding - Engine grows a stepwise move API (clearStep / finishMove; applySwap drains the same loop) so the animating web host and the one-call training path are the same rules by construction - C#<->TS parity checksum re-verified unchanged (78377593). - crazy-fruits-game.ts: animating host recording swap -> pop -> fall [-> reshuffle] timelines from the engine; view-only fall geometry. - crazy-fruits-render.ts: fruit-stall theme (bunting, strawberry-red wordmark, vector fruit art), canvas HUD (score/moves/best drawn on-canvas - loop runs outside zone). - crazy-fruits.ts/.html/.scss: unified Pointer Events - ONE path for touch (touchstart/move/end) and mouse (mousedown/move/up): drag-swap past a 0.35-cell threshold + tap-select-tap-swap; setPointerCapture; touch-action: none. Watch-AI scaffolding: tier picker (Random/Greedy/Expectimax/net), director driving the engine's scripted baselines, net loader (falls back to expectimax until the M49.3 ckpt ships); screen wake lock during watch. - Registered: route /crazyfruits + nav link + home card; ARCHITECTURE.md env + client-side tables. - tsc --noEmit clean. Gate note: live desktop+smartphone verification pending the user-run host (down at implementation time) - to be evidenced before the PR. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…pisodes) Three-run story (all 150k moves, masked Double+Dueling DQN, one lever per intervention per the no-thrash rule): 1. gamma=0.99 (the Snake recipe verbatim): loss exploded 0.62->8.7, eval regressed below random mid-run; on the 500-episode gate the keep-best net was +1.9% (OVERLAPPING) - the documented match-3 trap: the controllable signal is almost entirely immediate, and deep bootstrapping amplifies refill-cascade noise. FAIL. 2. gamma=0 (Q = expected immediate points incl. stochastic cascades): stable loss (~0.18), +7.8% CI-SEPARATED, above greedy - real learning, still under +30%. FAIL. 3. gamma=0 + the PRD s6 Risk-1 pre-registered mitigation - observation grows two per-action feature planes (immediate score, deterministic cascade value, /100; 448 -> 672 floats): mean 3552.5 +/- 83.3 = +57.2% over random CI-SEPARATED (gate >= +30%), +48.8% over greedy; expectimax-1 (4270.9) remains the scripted ceiling = future-training headroom. Ships wwwroot/models/crazyfruits.dqn.ckpt (LFS, keep-best from run 3, eval 3392.5 at 125k). Observation change verified: 22 tests green (incl. per-action-plane pins), C#<->TS parity checksum unchanged (78377593). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- CrazyFruitsNetParityTests: byte-level ckpt parse (the line-for-line reference for crazyfruits-net.ts) -> PgCfDuelingNet forward matches the SDK DuelingQNet within f32 tolerance through real checkpoint bytes; masked netAction returns legal actions. - Node simulation of the exact browser path (shipped crazyfruits.dqn.ckpt -> TS parser -> generated net on the generated engine): all four watch tiers play full legal 30-move episodes; strength ordering reproduces (random 2379 / greedy 2520 / net 3393 / expectimax 4507 over 20 eps). - PRD + PLAN M49 updated with all measured gate results; M49.4/M49.5 marked "built, live in-browser device check pending" (the user-run dev host was down throughout; headless evidence recorded in the PRD). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
LIVE Playwright against the running host (desktop 960px mouse + emulated phone 390px with real touch events): - tap-tap select ring lands on the tapped cell on both form factors; a human tap-tap swap cleared a 3-line (score 30, move 1, screenshot-verified) - mouse drag and CDP touch-drag both fire the gesture; illegal picks revert without consuming a move; window.scrollY unchanged through the touch drag (touch-action: none holds); zero console/page errors - watch mode verified live mid-cascade (trained net, score 330 at move 7/30, "+30" pop on screen); all four tiers exercised on desktop and mobile Screenshot-driven UX fix: a failed tap-tap swap now clears the selection like a successful one (the persistent ring was both a test false-positive and off-feel). PRD + PLAN M49 marked fully SHIPPED. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Member
Author
|
Live browser gate completed (host started per the clarified rule — port was free): desktop mouse + emulated smartphone touch both verified with Playwright. Human tap-tap swap cleared a 3-line (score 30 · move 1), drag gestures fire on both input types, illegal swaps revert without consuming a move, |
Replaces the hand-drawn vector fruits with fruit-cake-art.ts (cached 256px offscreen renders, blitted scaled - the per-frame cost is a textured quad). Six visually distinct picks from the 11-tier catalog, no two share a color: strawberry, grape, dekopon (orange), pear, pineapple, watermelon. Radius bumped 0.36 -> 0.42 cell (the art fills its circle), selection ring to 0.47. C# viz labels synced to the new fruit identities (labels only - observation unchanged). Verified live: board + watch-mode cascades render with the new art, no errors. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…+ PLAN entry 3-agent investigation (canonical Candy Crush mechanics, line-referenced engine impact, AI/training impact) -> docs/prd/CRAZY_FRUITS_SPECIALS_PRD.md + PLAN M50. Key locks: fully deterministic rules (zero RNG - passive bomb clears the most- frequent type, (r+c)%2 conversion orientations, lowest-run-cell cascade spawns; striped blast PARALLEL to the match, the "perpendicular" wiki wording is sprite paint); packed base-16 cell encoding (plain fruit keep 1..6); bounded activation worklist with the wrapped double-explosion riding the grid as an internal armed kind (stepwise animation API unchanged); swapIsLegal supersedes swap-must-match (action space stays 112); obs 672->928 => from-scratch retrain; v1 keeps gamma=0 with the per-action deterministic-value feature extended to price create+fire+ combos (PBRS at gamma=0 is a provable no-op), ONE pre-registered escalation (gamma=0.5 + 3-step + PBRS) gated by a new expectimax-2 baseline; game-over = 30-move rounds (owner decision - measured zero deadlocks ever, so game-over-on- deadlock would never fire; 30 keeps round scores comparable to the trained tiers). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Packed base-16 cells (kind*16+type; plain fruit keep 1..6); typewise match scan now records runs; creation resolver (bomb > wrapped > striped, blast PARALLEL to the match, swap-cell/lowest-cell/intersection spawns); bounded activation worklist with chains; wrapped double explosion rides the grid as an internal armed kind (stepwise animation API unchanged); stageSwap + swapIsLegal (bomb/special+special always legal, action space stays 112); all six combos; passive bomb zaps the most-frequent type (zero-RNG lock); per-move creation/fired telemetry. - Fire-only scoring (owner decision): creations score nothing in-game; the training env gains ShapeCreationRewards (+40/+60/+100 to the REWARD only, train env only). - Planning (immediateScore/deterministicValue) runs REAL clearSteps and restores board+context byte-identically (purity test: no RNG, WriteState bytes unchanged). - Observation 672 -> 928 (+4 kind planes, /300 feature planes); colored specials set both planes; bombs colorless. - Gate GREEN: 42 tests - every creation shape (incl. priority + cascade spawn), every activation (wrapped two-step timeline, never-armed-on-stable invariant), every combo hand-scored exactly (bomb+bomb 640, striped+wrapped 390, ...), chain reactions, invariant sweep, legality brute-force; C#<->TS parity RE-PINNED at 533753109 via the NOW-COMMITTED tools/cf_parity.mjs (episode exercises creation, chains, combos and a mid-episode reshuffle byte-identically). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- RewardScale 30 -> 100 (random means ~86 pts/move with auto-firing specials); per-action feature planes /300; ShapeCreationRewards wired to the TRAIN env only (--no-shape opts out); stale-width ckpt guard in --baselines. - New single-source tiers: expectimax-2 (beam-8 two-ply deterministic - the baseline that can see create->fire) and specials-greedy (immediate + creation weights); per-tier specials created/fired stats in the gate table. - Gate GREEN (500 seeded episodes): random 2612.2+/-67.5, greedy 3446.0+/-99.6 (+32% - specials widened the skill gap from M49s +6%), specials-greedy 3850.4+/-113.8, expectimax-1 5841.2+/-159.8 (+124%), expectimax-2 8368.8+/-177.8 (+220%). Env validation: random = 31% of expectimax-2 (< 70%). Expectimax-2 gap +43.3% over expectimax-1 => the M50.3 escalation trigger is armed. Directed tier tests: greedy provably takes the bomb swap; specials-greedy provably builds. 45 CrazyFruits tests green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…orm-then-trigger 1. Striped blast is PERPENDICULAR to the creating match (real Candy Crush rule: horizontal 4-match -> vertically-striped fruit clearing its COLUMN; paint shows the blast). Overrules the research agents parallel reading. 2. Combo blasts centre on the gesture LAST-SELECTED cell: stageSwap(action, targetCell); web host passes the second tap / dragged-to cell; AI + baselines default deterministically to the action bottom/right cell. Directed test pins that the same swap staged toward each end fires the matching column. 3. Specials FORM before the step activations run: a fresh special lands on its spawn cell first (unmarked - survives untouched), and fires IMMEDIATELY if any blast/combo of the same step reaches it. Directed 190-point double-match test: a fresh striped triggered by a wrapped box within its own creation step. 47 tests green; parity re-pinned 533753109 -> 801202210 -> 995400597 (score 95550) via tools/cf_parity.mjs; 500-episode baselines re-measured on the final rules (random 2598.7 / greedy 3497.9 / specials-greedy 3867.1 / expectimax-1 5931.4 / expectimax-2 8135.0; env validation 32% < 70%; e2 gap +37.2% keeps the escalation trigger armed); the in-flight training run was discarded and restarted per change. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ts, 30-move rounds - Square candy wrapper (owner-requested): translucent gold rounded square with folded corner tabs behind the fruit + gloss band and double rim in front - the fruit stays visible inside. Stripes: thin (0.1r) white bars with dark outlines painted ALONG the blast axis (readable on light fruit); colorless sugar-bomb sphere with sprinkles. - Pop step enriched from the engines lastClearedBy/lastCreated: striped beam flashes, blast rings, bomb zap glows, creation sparkle-ins. - stageSwap gesture wiring (combo centre = last-selected cell); six watch tiers (+ Specials-greedy, Expectimax-2); 30-move rounds with the round-over screen (score / best / measured tier bars / tap to play again); deadlock still reshuffles mid-round; home card + copy updated. - Gate GREEN: tsc clean; headless smoke (60 greedy moves, engine-exact grids, specials firing - 8130 pts vs ~4360 pre-specials); LIVE Playwright: watch tiers screenshot the square-wrapped pineapple / striped fruit / two bombs, and a real 30-move human round (164 pixel-diff-probed swaps) ended on the round-over screen, tap restarting a fresh round; zero console errors. M50.3 training deferred by owner call - runs in background, ckpt + gate table land as a follow-up; the net tier falls back to expectimax until then (stale-width guard). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Member
Author
|
M50 (Candy-Crush specials) is now in this PR — striped / wrapped / sugar bomb + all six combo swaps, single-sourced in the |
- "Endless mode" toggle (human play): keeps playing past the 30-move round and dismisses an already-shown round-over screen, resuming the same board. Any game endless touches is permanently exempt from the "best" indicator (bestExempt latch per game); rounds + best resume with the next normal game. HUD status shows "endless - score wont count for best". - Striped overlay bars thinned 0.22r -> 0.10r (less than half), per owner. - Verified: tsc clean; headless semantics smoke (round closes at 30 / endless never closes / resume-from-round-over / best frozen once endless touches / rounds+best resume when off); live toggle screenshot, zero console errors. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…nctionality tests - StepwiseHostProtocol_IsByteIdenticalTo_ApplySwap: the web pages exact engine calls (stageSwap -> clearStep/collapse loop -> finishMove) vs the atomic applySwap, 10 seeds x 30 mixed greedy/random moves incl. specials and combos - identical grid, score, RNG stream, reshuffles and fired-telemetry every move. The browsers game functionality IS the tested engine functionality by construction (TS equality separately pinned by the parity checksum). - tools/cf_host_tests.mjs (committed, `node tools/cf_host_tests.mjs`): the host layers own rules - round closes at exactly 30 (not 29), round-over blocks moves, endless dismisses the screen and resumes the same board past 30, best freezes the moment endless touches a game (incl. games started under endless), rounds + best resume when toggled off, illegal swaps never consume a move. Runs the real crazy-fruits-game.ts under node type stripping (temp-copy import rewrite); no rendering involved. 48 CrazyFruits xunit tests green + both node harnesses (parity, host) green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…er Q) - PassiveBomb test extended: after the move, the striped that set the bomb off AND the bomb itself are off the board. - New Wrapped_TriggersBomb_AndBothAreRemovedByMoveEnd: a wrapped 3x3 catches a bomb; exact step-0 score computed dynamically (match + box + most-frequent-type zap); by move end the wrapped (via its armed second explosion), the bomb, and any armed shell are gone. - Lab --baselines: the net row now reports specials created/fired per episode (needed for the M50.3 usage gate). 49 CrazyFruits tests green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…fallback) The gamma=0 run passed the +30% floor decisively (+53.9%, 4000.5 +/- 128.8 on 500 episodes) but missed the 64% gap-share bar (42%) and the 2x-random creation-usage bar (5.7 vs 7.3/ep) - exactly the hold-for-combo blind spot the PRD pre-registered the escalation for (expectimax-2 gap +37.2% >> 10% arms the trigger). - CrazyFruitsDqnOptions : DqnScoreOptions adds NStep (campaign feeds it to the trainer; plain options = defaults). - CrazyFruitsEnv.ShapeSpecialsPotential: potential-based shaping, Phi(s) = sum of on-board special option values (40/60/100), reward += gamma*Phi(s2) - Phi(s1) - policy-invariant at the matching gamma, prices HOLDING a special without rewarding hoarding (provable no-op at gamma=0, hence escalation-only). - Lab: --nstep, --pbrs (escalation config = --gamma 0.5 --nstep 3 --pbrs; PBRS replaces the creation bonus on the train env; eval env stays bare). 49 tests green. Escalation run launched; ONE attempt per the no-thrash stop-loss. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…est counts PRD/PLAN now carry the full honest M50.3 state ahead of the closing commit: attempt 1 (gamma=0) 4000.5 +/- 128.8 = +53.9% over random (gate 1 PASS, +14.4% over greedy) but 42% gap share and 5.7 created/ep miss gates 2/3; the ONE pre-registered escalation (gamma=0.5 + 3-step + PBRS) is running. PLAN lock text gains the third correction (form-then-trigger), fire-only scoring and the endless toggle; test counts synced to 49 + the two committed node harnesses. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…to the nearest plain cell Owner bug report: dragging a striped fruit into a line of 4 neither fired its blast nor left a new striped elsewhere in the line. placeCreation overwrote-and-unmarked whatever held the spawn cell, silently deleting the existing special on all four creation paths (striped swap-cell, wrapped pivot, bomb match-5, cascade lowest-cell). Rule (matches real Candy Crush): a spawn cell already holding a special is never overwritten - that special stays match-marked and fires through the existing form-then-trigger pass, and the creation relocates to the nearest plain cell of the shape (outward along both arms from a wrapped pivot; ties toward the lower flat index). A shape with no plain cell creates nothing - every special in it simply fires. - 8 directed tests (reported drag exact-scored at 180; wrapped variant + armed refire; plain-spawn regression guard; all-special run, zero creations; equidistant tie-break; bomb relocation with priority intact; wrapped-pivot; cascade relocation) - suite 57/57 green - Parity re-pinned 563660409 (C# = TS via tools/cf_parity.mjs); host harness green - Baselines re-measured on fixed rules (M50.3 bars updated): random 2613.9 greedy 3499.2 - e1 5974.6 - e2 8139.1; validation 32%; e2 gap +36.2% keeps the escalation armed; the in-flight escalation run was restarted on the fixed rules Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The page now carries a tribute paragraph: Crazy Fruits came from KidCity.be, the Belgian Flash-era children's portal; both are gone today. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…er keeps the promised special Owner report round 2: a wrapped dragged into a line of 4 still never left a new striped standing. Two-agent audit showed M50.5 was functionally correct (the dragged wrapped fired and the striped WAS created, all 13 geometries) - but the wrapped's own 3x3 always covers the nearest-plain relocation cell, so the relocated striped chain-fired in the same step and the armed second explosion re-covered it a step later. A dragged striped only spared it when the blast axis missed - exactly why striped drags looked fixed. Rule: a creation displaced off its preferred spawn cell is shielded from blasts (beams, boxes, armed refires, bomb zaps, combo areas) for the rest of the move - the fired special cannot consume its own replacement. Matches in later cascade steps still consume it normally, and creations placed on their preferred plain cell keep the form-then-trigger chain rule (the 190-point test is untouched). - shielded[] list, set only on relocation, skipped by markCell (the single blast-marking entry point), cleared by stageSwap - never outlives a move - collision tests updated to shield semantics: striped drag exact-scores 100 with the fresh striped surviving; wrapped drag 90 surviving BOTH explosions; wrapped-pivot 80 keeping an unarmed wrapped - 57/57 green - parity re-pinned 481681208 (score 95950, up from 86340: surviving specials fire later - the shield is score-positive), C# = TS verified; host harness green - the completed cf7train run (trained without the shield) is void; training restarts from scratch on the final rules Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…an and lost Final-rules gate eval (500 eps, seeds 5000+e; bars: >=3392.0 / >=4747.8 gap-share / created >=7.3 & fired >=5.6): - SHIPPED cf5train (gamma=0 + creation-shaping): 4040.4 +- 128.9 = +54.9% over random, CI-separated (gate 1 PASS, +15.1% over greedy); gap share 43% (gate 2 MISS); created 5.81 (MISS) / fired 5.75 (pass). The net transfers unchanged across the rule fixes because its per-action feature planes are computed by the live engine at inference. - The ONE pre-registered escalation (gamma=0.5 + 3-step + PBRS) got its clean from-scratch run on the final rules (cf8train, 400k): 3408.2 = +30.6%, gap share 24%, created 4.66 - worse on every gate. Bootstrapping loses to gamma=0 on match-3 refill noise even with PBRS; the M49 lesson reconfirmed. Two earlier escalation runs were voided by the in-flight rule fixes (cf6train killed at 210k; cf7train trained without the shield, only +33.4% under final rules). - Stop-loss invoked per protocol: best net ships, gates 2/3(created) are reported missed, no further training. Hold-for-combo (the +36% e2 gap) stays unclaimed by reactive nets; the recorded future lever is search-guided play, not another reward schedule. Round-over screen gains the net's challenge bar (~4 000). PRD/PLAN updated with the full verdict and final-rules baselines. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
M49 — Crazy Fruits (match-3) + primitive net
New game + AI in the KidCity (kidcity.be) Crazy Fruits spirit: swap two adjacent fruits to line up 3+, cascades chain, works with mouse and touch. PRD:
docs/prd/CRAZY_FRUITS_PRD.md(planned via a 4-agent investigation; the original Flash SWF is unrecoverable, so this ships assumed-standard match-3 rules under the confirmed fruit-stall theme).What's in it
crazyfruits_solver.pgsingle-source (C# + TS twins): rules, proportional cascade scoring (10·(k+1)/fruit, +20/+50 line bonuses), hard legal-move mask (match-producing swaps only), deadlock reshuffle in-engine (the mask never goes all-false), Schrage-minstd RNG (exact i32 both sides), 672-float observation, scripted random/greedy/expectimax-1 baselines, dueling-net forward, and a stepwise move API for the animating web host. Gate: 16 unit tests + a seeded 1,000-move episode byte-identical C#↔TS (per-move full-grid checksum 78377593).CrazyFruitsEnv(masked, stateful, truncation-only) on the M46DqnScoreCampaignspine;--game crazyfruits(+--baselines Ngate table). Gate: 500 seeded episodes — random 2259.7±49.9, greedy 2387.0±49.3 (CI-separated), expectimax-1 4270.9±98.3. Finding: cascade planning is the skill (+89%); line size is nearly irrelevant (+6%).wwwroot/models/crazyfruits.dqn.ckpt(LFS). Expectimax-1 (4270.9) is the future-training headroom./crazyfruits): canvas fruit-stall renderer (bunting + vector fruit art), animations driven by the engine's stepwise API, unified Pointer Events (one path covering touchstart/move/end and mousedown/move/up): drag-swap + tap-tap-select,touch-action: none,setPointerCapture.crazyfruits-net.tsckpt parser (byte-level reference:CrazyFruitsNetParityTests).Verification
dotnet run --project src/RLDemo.Weband visit/crazyfruitsto complete it. All headless evidence is recorded in the PRD.🤖 Generated with Claude Code