MDK2 pending Medium Enter: finite audio refill and observation cost

The accepted-runtime history stops during buffered audio decoding, before a witnessed Medium press. There is measurable additional observer work, but no proved emulator/driver scheduling defect or safe causal remedy for this exact history. Production files are unchanged. This finding does not authorize a retry.

Inputs are published 337356d17278bf2a5df8726437d2dabea3933e4d plus a7d8905b1e5495a17354da7f94736bddf6e5d781, with captures in scratch/runs/20261104-mdk2-accepted-runtime-binding/. Evidence for this investigation is scratch/runs/20261105-mdk2-pending-enter-decoder-cause/ (RUN). The earlier decoder-return, GC/callback, keyup/feedback, and timer/deferred-load findings remain distinct histories and static contracts; they are not backfilled into this run.

The offline audit verifies the original EXE SHA b3e0226e411abff17ddabb723e493e95579490cb49bfe3787b1eade1144e27c1, ten instruction contracts, all 28 trace-end append prefixes and matching main/file sidecars. The final history has 465000 boundaries, 29 samples, three output contexts and two decoder returns. At every post-step sidecar, main is thread1 under saved returns 47ca7e, 47a928, 47b393; file thread2 is independently at the idle-file-poll boundary. The completed pointer resource remains unchanged. The trace is continuous, but its selected records and sparse PCs are not an exhaustive record of every executed instruction.

Direct evidence Meaning
Output routine 47d980, call 47ca79, checked return 47ca7e at boundaries179389 and335388; EAX32768, ESP=frame+28 Two finite 32768-byte requests returned.
Object remainingWords11410842 then11394458; decodedWords0 at both returns 16384 words were consumed between these returns. Zero buffered decoded words is not stream EOF.
Final PC47deb8, decoder frame ESP074ffcd8, object0847a800; stack produced24576/requested32768 The third observed request still has8192 output bytes unproduced. This is progress through transform/output work, not a wait on queued Enter.
Outer 47b390: call [ebx+8], then 47b549..47b552 decrements [esp+24h] and loops to47b331 while nonzero A decoder request return is inside an audio refill loop, not necessarily a main-loop return or input poll.
Main 402171 calls maintenance451540; after return, 40217c compares the periodic counter with5 and conditionally calls GC443070 at402187 Periodic GC can be another later dependency. There is no GC entry/exit record in this accepted window. The earlier run's GC endpoint is not this run's cause.

The original call-chain interpretation remains 402171 → 451540 → 47c600/47c610 → 47bf90 → 47b1a0 → stream callback → 47a880 → 47ca50 → 47d980. The three inner saved returns are directly present in this history. The deeper main-loop chain comes from original bytes and the earlier bounded stack, not a new capture of the stream's callback pointer. DirectSound Lock/Unlock-shaped vtable calls surround the refill, at47b2a1/47b572; its stream object and remaining loop count were not captured here. No full-stream decoding prerequisite for future input is established.

At4096, the sidecar covers [074ffc78,074ffd38). Reconstructing the original stack adjustments puts the outer loop counter at074ffd64, 44 bytes beyond the captured end. Consequently neither the entire stream's remainingWords nor the third request's8192 bytes gives the number of remaining refill iterations. The exact immediate control dependency is the requested-byte comparison at 47d9e4/47d9e8, followed by the caller and outer loop counter at47b549. The missing counter cannot be recovered by interpreting uncaptured bytes as zero.

Enter has three separate states. The frozen renderer retains WM_KEYDOWN13 and WM_CHAR13; the keydown carries asyncPressPublished:true. The original renderer-input.js publishes asynchronous press state when enqueueing normal keydown, and host.js can publish the live key bitmap to the Worker without dequeueing these messages. Thus queue presence does not imply async state was unpublished, and the stamp does not prove a current key-bitmap read or native poll. host.check_input → renderer.takeInput performs message delivery; the game's 46cee0 → 46cf30 → 454cd0 command sweep is another operation. At the final capture, lastInput and activeInput are null, command62 has raw0/previous0/active0, current=diff, scheme=menu, showmenu1. No consumed Medium press, keyup, callback, full-rest-after-release, timer or new-level witness follows from an ACK or that enqueue stamp. A native async poll could also occur without draining the message queue; neither is invented here.

The aggregate menu allowance is already512:16 observation +48 New Game +448 Medium steps. The extra32 (maximum544) is conditional on the existing released Medium path and cannot fund this pending press. Every post-step trace-end reports the actual received budget1000 and ranBlocks1000. The first16-step interval accounts for33000 boundaries; the following27 each account for16000. There is no evidence that review/checkpoint commands secretly advanced frozen steps. There is evidence that the selected native work per step is small.

The current observer rechecks35 instruction pins totalling502 bytes at every native boundary, including bytewise guest_to_wasm calls and hex-string conversion. RUN/audit.cjs calls the unmodified observer for1000 boundaries against an explicit getter-only memory fixture and counts534 translations per boundary. The exact predecessor observer from12a87f179 (SHA32a03239…) checks9 pins/122 bytes and makes154 translations in the same fixture. The published timer observer399d2128d (SHAdd37b5ba…) adds380 such translations per boundary:176700000 additional translations over465000 boundaries, for this matched fixture path. Actual resource/special-boundary reads can add work; this is an operation count, not a timing measurement or whole-history replay.

This observation work runs inside the Worker's timed slice. The exact accepted host feedback statements use round((ranBlocks*12/ranMs)/1000)*1000, clamped to1000..configuredSteps, after sufficiently spent non-input-burst slices. An isolated fixture executes those original statements:1000 blocks at1/6/12/100ms select12000/2000/1000/1000; input-burst and trapped-slice negatives leave the prior budget unchanged. The earlier native discriminator already proved the observation/quiet-trace cost difference; its long decoder is deliberately not rerun. The accepted bridge already suppresses quiet block-log RPCs. Reapplying that old bridge fix cannot remove the present pin-check work.

The added work is therefore a concrete way observation consumes wall time and can reduce native progress within the fixed step allowance. Whether every check is necessary at every boundary is a separate equivalence question: simply memoizing pins, checking only at samples, subtracting observer time from feedback, disabling tracing, or changing budgets would alter existing byte-mutation, deadline, ownership or scheduling behavior. No independently proved correction with the required negatives and complete parity is available from this evidence. The captured nativeSlice records lack per-slice ranMs and the host adaptive/burst state, so they cannot assign the observed1000 solely to observer cost or prove an optimization would reach consumption within512. No production edit is made.

The remaining discriminating evidence is specific: the live outer stream's callback/object fields and counter[outer ESP+24h], its buffer/cursor/lock state, and the decoder's remaining callback/input state; then same-history maintenance return, any periodic GC return, native command update and queue/async-key state. For a scheduling correction, independent short fixtures must establish unchanged wrong-owner/stale/byte-change/bounds/failure/queued-input rejection and both-module full-memory/GPR/command parity, alongside separated observation timing. Neither the missing stream state nor the input boundary can be reconstructed from the 16-byte decoder windows. This is the concrete blocker to a causal remedy, not a request for more resources or authorization for new original execution.

Validation: four existing focused native tests pass. Both accepted module/source pairs execute the original small native fragments against controls with complete memory hashes, registers, commands and trap parity. Separately, both retained module variants pass the four-word decoder output/replay, register/lazy-flag and output parity, quiet/forward controls and import-failure parity; these are explicit fixture data, not this stream. The ten-contract/prefix/suffix/work-count/feedback audit passes. Its initial epilogue-address transcription failure is retained in audit-initial.log; corrected47d9ee is byte-verified. No full build is claimed for this documentation-only change, and no module was mutated or restored. The accepted native tests read and assert the pinned module hashes before execution. A later audit of the shared build paths sees the legacy worker's new63a43aa7…/09a64727… modules; those are recorded as later path observations, not substituted for the tested accepted pair. The retained decoder fixtures use their existing historical paths. No shared module is copied or rolled back.

All50M/360s owner/record/byte/input/ACK, active0 full-rest, frozen review,16ms and 512/544 caps, original assets/code/config remain. No Chrome, game, boat, heavy runtime lock, full long decoder, subworker, deployment or HeroesII ran. Git SHA plus relevant binary/new-file patches preserve source; media/modules remain existing-path SHA references. Box ID/state/expiry: NA; coordinator owns shutdown.