Mad Bomber: authenticated browser logical FPS

94 logical gameplay loop entries / 4,438.655 ms = 21.1775864536 FPS on exact main f09e09c6643997986bbafad69c37adae4906df90. This is one short, actual instrumented wall-clock interval, not display FPS, a sustained average, or an uninstrumented performance estimate. One browser session completed in 14.204 seconds, under its 60-second deadline; no second session was needed.

Game-specific semantics

Original madbomber.exe SHA256 72ad26d11ddabf72c668b69eda0c7c1daa634f5cade631303ba4e05eb5f5cc09 is UPX packed. unpack.py transcribes its own decompression/filter stub at 0x416600..0x416706 into a read-only analysis copy. It does not change the launched EXE. The stub, decoded bytes, recovered import names and disassembly are preserved. The entire 6,848-byte live game-loop range 0x403d01..0x4057c0 matches the independently decoded original, before sampling and at every sample. This live comparison is the check on the analysis decoder, not an assumption that an unpacking script must be correct.

One entry at 0x403d01 is one iteration of the original gameplay loop:

• 0x403d0b calls SDL_GetTicks via 0x40a980; 0x403d13 increments the game's frame variable at 0x40fd40 once. SDL_PollEvent is the inner input loop at 0x403d20, so multiple input events do not create extra frames.
• Active-play state at 0x411a40 gates bomber movement and bomb emission at 0x404197 and 0x4042fa. The bomb loop at 0x404687 updates each active bomb's Y at 0x4046ed. Bucket and drawing work are inside this same outer loop; SDL dirty-rectangle updates are not counted as frames.
• 0x405788..0x4057ab performs the original SDL tick/delay check, then 0x4057bc branches back to 0x403d01 unless the game exits this loop.
• The game's frame variable resets at initialization (0x403c8b) and the fire/start transition (0x404039). It is useful corroboration, not the authoritative reset-free measurement counter. During the measured interval it advances 27 → 121, exactly matching marker 53 → 147. The offset is 26 in every one of the 46 samples.

The existing WASM set_logical_frame(0x403d01, 0) installs the decoder's logical-entry marker. Pace 0 counts without changing guest pacing; get_logical_frame_paced() remains 0. The exact source has one counter increment site, no reset writer, and per-instance initialization at zero. Setting a new address flushes decoded code at a block boundary; that cost and subsequent marker/cache effects are not claimed to be free. The flush precedes both the short probe and primary interval. Decoder and micro-op compiler rules prevent a fold or compiled loop from swallowing the marker.

These addresses belong to this original Mad Bomber build and were derived here. No Hype address, renderer swap, requestAnimationFrame count, expected 30 FPS target or guest clock is used to calculate the result.

Live owner, short gate and rejection controls

The registered route honors threads:false. Live inventory finds a main instance with exported thread ID 1, plus one cooperative instance with handle 921600, internal tid 1, exported ID 2. Both execute in the page clock domain; there is no guest Worker backend. The main instance equals threadManager.mainInstance. Both counters were enabled and sampled; only main emits this game-loop marker. The cooperative instance stays at 0.

Before the primary interval, nine genuine timestamp/count pairs over 718.86 ms show 15 entries on main and zero on the cooperative instance. The executed task-specific validator rejects the live wrong owner and six mutated copies of this probe: disabled marker, changed owner ID, changed clock origin, decreased count/reset, lifecycle mutation and inactive gameplay. All seven rejection results have FPS null. The gate was saved at 17:35:31.351Z, before primary sampling began at 17:35:31.544Z. validate.js is a new capture-local contract checker; no existing production validator or Hype validator was changed or relabelled.

Every sample records performance.now() paired with each instance's counter, clock origin, current thread ID, enabled address/pace, paced count, identity, live code match, running/active state, owner count and sampling cost. Forwarding wrappers record host logical-setting calls and cooperative map lifecycle changes; an instance-property observer records main-instance replacement. They preserve original calls/results. All event records remain empty. Original address/pace values (0/0 on both instances) were restored, timers and wrappers removed, and the owned browser/server closed normally. The preserved source accounts for the counter's reset-free lifetime; samples alone are not claimed to exclude every hypothetical hidden reset or unaudited script.

Actual interval and environment

The measured page timestamps bracket 17:35:31.544–17:35:35.983 UTC. Duration uses the precise monotonic timestamps, not rounded UTC strings. There are 46 paired samples. Ordinary Enter selects One Player, Space starts bombs, and DOM mouse events move the buckets. Personally inspected measure-before.png, measure-during-right.png and measure-after.png: active falling bombs, bucket movement and score 0 → 1 → 3 bracket and identify this interval. All active-state samples equal 1. No restart, pause, extra fire or menu action occurs during it. The result does not extend to the later Press Fire state or to full-game completion.

Chrome 151.0.7922.108, Node 24.18.1, viewport 1000×760, game client 640×480, assigned 4-vCPU/8-GB Boat bx_ppatp9cq. Exact route: http://127.0.0.1:8080/index.html?app=madbomber&debug, through the unchanged registry and browser shell. Live renderer is Win98Renderer with CanvasRenderingContext2D. CDP reports ANGLE/SwiftShader (SwANGLE 5.0.0), a software graphics environment, not hardware GPU qualification. The toolbar's WebGL selector is not evidence that this SDL game used a WebGL renderer. Display FPS, hardware-GPU performance, audio quality, full completion and Windows parity remain unqualified.

Sampling both owners plus authenticating the live code costs 24.925 ms total, median 0.5075 ms, maximum 1.210 ms, or 0.56154% of the measured interval. Maximum timestamp/count read span is 0.015 ms. The one in-interval screenshot costs 234.872 ms of driver elapsed time; gameplay continues and that time is included. There is no subtraction, causal overhead baseline or claim that this isolates marker/cache, event-loop, CDP/screenshot contention or timer costs. overhead.json preserves details.

Provenance and diagnostics

No rebuild: the retained canonical module was inventoried and verified as 728e1de80ec20a284fc41b339895ee8ac29f16e4ded83325b27576cfacd4466c, 1,878,508 bytes. Its actual Chrome response body has that same SHA256 and size at /build/wine-assembly.wasm?v=dev. Canonical build revision is 30a2c3fb1b5c2bfa8a6b001d7394d5cf50df295d; retained build exit is 0. Exact Git comparison to intended main changes only docs/evidence/status and two unrelated Army observer tool/test files, absent from this route. Browser, compiler, build inputs, package lock and production sources are unchanged. The 7,299-entry canonical and 6,909-entry staged manifests match their Git trees; all staged files match before and after the run.

All 168 asset pins and 2,407 dependency files verify pre/post. The original retained archive is unchanged: SHA256 b5353b73674a1af6109cb9bf3af5f7d0383308b7ef0b47d52139813ba3ee98ef, 3,069,702 bytes. It was transferred to the fresh box without redownload or local duplication. Exact-lockfile npm ci exit 0. No production source, guest memory, API behavior, clock or original binary was modified.

Full diagnostics are retained: 109 console messages below the 2,000 cap, 139 network records, four failed requests and four optional 404s (build-info, icon, favicon twice); page errors are empty. This run has one console warn (the deprecated mobile-meta warning). Audio was not investigated. Harness, Chrome close and server cleanup complete normally; session process exit 0. The evidence checker and offline verifier results are retained separately.

Evidence and handoff

Bundle: scratch/runs/20261021-madbomber-browser-fps/. verify.js replays identity, recovery seals, live-code/owner authentication, seven controls, measurement, matching guest frame deltas and cleanup. Four screenshots and all diagnostics are recovered individually, with 24 remote-file seals (1,643,807 bytes), rather than making another archive. RECOVERY.md preserves staging/reproduction instructions; publication receipt records the task commit and exact remote branch. No main merge or deployment.

SAME bx_ppatp9cq is left RUNNING, API state idle, health ok, expiry 2026-10-21T19:28:40.555Z. Owned harness PID 26854, Chrome 26876 and server 26865 are absent; port 8080 is closed. Complete staged source, dependencies, originals, canonical module and evidence remain at /home/user/madbomber-fps. Coordinator owns snapshot-stop and lifecycle. Handoff is well before 19:05 UTC. No workers, new boxes, resumes, box SSH, public hosting, root task/status edits, shared Git configuration, laptop/Heroes II/Warcraft III or fleet changes.