Stronghold native copy and publication observation transaction

The missing transaction from published a3dfef099b66941f488f2420d1042a2dd0b9812c is implemented: the owning native cycle, actual selected primary fallback copy, destination lifetime, ordinary compositor publication, ACK and capture identity now reach freshShot and the route's diagnostic consumer. A complete receipt can describe setup in progress. It cannot accept pixels, admit input, or claim gameplay. Those three flags remain false, including after a checked whole-cycle return and during retained-pointer partial writes.

This assignment executed synthetic native fragments and real host composition on both shipped modules. It did not execute an original application or browser, retry MDK/Stronghold, or qualify actual setup-to-gameplay behavior.

Release gate and source identity

Work started against main d31a2808d4a262d9f7b611f8c13a437ae0ebd605. Source preparation stayed in the assigned ignored run until the ordered D3DIM 03b16e2553ddbe7d879fea8524e9df479a97fa07 and MDK a80cc075bf917ad54610130a98f46263a1590b4e final releases. The D3DIM row's actual wording, FINAL RELEASE with explicit files and LOCALHEAVY runtime ALLBUILDSFINISHED, is retained verbatim rather than normalized into invented text. MDK published FINAL runtime AND file RELEASE ... ALLBUILDSFINISHED.

implementation-release-gate.json records the independent 06:56:59Z gate: five D3DIM browsers and the MDK browser had actual normal exits, exact PID/start identities were absent, six profiles were absent, thirteen ports refused connections, both build/runtime locks were free, and disk exceeded 1.5 GB. Shared source claims and edits followed that gate. No predecessor files were edited while owned. The predecessor MDK reason-5 fault and lost terminal Worker journal remain unknown; this work supplies no historical invalidator claim.

Evidence is in scratch/runs/20261105-stronghold-native-copy-publication/. Git plus the relevant binary diff and new-file patches preserve source identity. Modules and media are referenced by existing paths and SHA-256, without source snapshots, archives, module copies, or repeated game media. Final test/process, module/input hashes, cleanup, and publication identities are recorded there. The carried transaction-contract.md and proposed-*.patch remain historical UNTESTED proposals; the implemented Git diff and executed results supersede them as evidence.

Native profile and history

The existing MDK arm ABI, profile-1 target set, validity reasons, depth bound, CALL/RET instrumentation, lifetime guards and caps remain intact. An explicit pending_context_arm_profile(epoch, 2, 0x400000) is available only once before execution. Invalid, late, repeated or foreign requests do not revive coverage. The Stronghold owning preload arms it before returning the original instance to Worker initialization, binding the exact Module/memory/endpoint, executable SHA, module SHA, Worker URL, full target/session epoch and native epoch.

Profile-2 target Observed role
4e95f0 Input manager admitted by the original gate
4e9180 Scene-manager pass
4292c0 Draw nested under that scene-manager pass
4e9400 Widget passes with distinct actual callers
4ea0a0 Separate post-widget work
467b90 Presentation pass

The observer retains the original 576b9a gate's four operands, EBX, manager and actual executed successors. It requires the checked input-manager return or the executed bypass successor before observing the 576c51 reset block and its actual successor/store result. Pass receipts include entry PC/ESP, caller PC/ESP, manager/head, native tracked frame and checked return PC/ESP. Draw and prior pass history must independently exist. A sampled pending frame, an empty ledger, hover, pending selection, or a fade value cannot substitute.

MDK pendingAbsence explicitly rejects profile 2. Conversely, a valid MDK profile-1 depth zero cannot supply Stronghold coverage. The MDK integration's exact observer-source seal is updated for this stricter consumer check; its immutable original-runtime resolver and historical runtime manifests are not amended. No original-run authorization is introduced.

Actual selected copy and ordinary publication

host.js forwards the existing dirty/retained-lock policy as observation data; its decisions remain unchanged. lib/host-imports.js emits a receipt only after the selected memory-to-destination copy. It records allocation generation, source slot/storage extent/geometry/flags, selection, owning memory/window, destination layer/canvas, clipping, copy generation, lock policy and frozen step. Startup census generation zero remains unknown. Slot reuse, even at the same address and size, invalidates the old receipt.

lib/renderer.js exposes passive hooks after actual transfers and successful ordinary publication. The observer follows the selected destination through windowed blits, exclusive intermediates and final filter publication. It also checks output/filter generation and guards destination writes and canvas size setters, including a same-size reset. Publication failure cannot reuse the previous receipt. No extra present, flush, readback, guest write, time change, scheduler action, lock-policy change or hold release is requested.

The transaction pins the live host, process, instance, memory, renderer/context, window/layer/canvas, broker, guest/link/Worker/Module, endpoint and Worker URL. Foreign or replaced lifetimes reject. Normal calls preserve their original receiver, return/Promise, exception and order; cleanup restores owned hooks. Unknown output paths stay unknown. Tests cover software Canvas2D output and the real software presentation filter; GPU/browser output is not qualified.

This establishes the actual copy-to-publication dataflow. It does not identify an original graphics operation that caused every source byte, prove complete pixel coverage/occlusion, or certify that a retained pointer finished its write. The older operation-to-upload relation remains unknown.

Capture and route consumer

createAttachment.capture returns the joined native/host capture transaction. The generated freshShot returns it as nativeObservation; the generated route adapter passes it to diagnosticScene. The consumer recomputes the join and checks the actual PNG digest plus capture name, step, command and epoch. The previous setup screenshot whitelist and exported hash list are removed.

The join requires complete bounded running reports, contiguous record identity, the executed gate/reset/pass/draw history and checked returns, current selected primary allocation/copy, the actual compositor lineage/publication, a complete 64-step ACK and its matching capture bracket. Native boundary count and native receipt must remain unchanged from ACK to capture. This detects execution during a screenshot even when PC and ESP end at the same values.

Missing, malformed, foreign, stale, replaced, partial, terminal, pending-async, or exhausted provenance returns unknown or a latched terminal rejection. The consumer distinguishes a proved in-progress setup diagnostic from completed cycle status, pixel acceptance, input readiness and gameplay. It grants none of the latter conditions.

Executed validation and limits

The focused test executes real native CALL/RET and original gate/reset/draw instruction windows with explicit synthetic data, helper bodies and clocks. It uses actual WineAssembly.getImports, dirty/retained-lock decisions, fallback imports, renderer, filter, production observer factories, attachment, generated freshShot, and generated route capture adapter on the same instance and memory. Only browser/CDP delivery and screenshot transport are substituted; PNG bytes come from ordinary fixture output. These are fixtures, not original Stronghold evidence.

On each shipped module it exercises:

Passing focused regressions are test-stronghold-native-copy-publication.js (final receipt focused-publish.json), the full test-stronghold-owning-callback-observer.js (observer-latest.json), test-stronghold-natural-setup-dependency.js (dependency-final.json), all 90 test-mdk2-native-pending-context.js checks (mdk-native-final.json), and test-stronghold-local-runtime.js --observation-contract (local-observation-final.json). The full flock /tmp/wine-build.lock bash tools/build.sh passed (build-publish.json, exit 0 without a signal). Each receipt records the actual process identity and exact absence after exit.

The historical multi-fragment draw fixture changes EIP/ESP between independent cases, which correctly invalidates continuous native proof. Its legacy draw/control consumer is explicitly isolated in test serialization; both arms retain the same native profile, and its 20 draw cases/30 negatives per module, 24 host structural negatives, and full memory/register/command parity pass. The separate new test exercises the complete unmodified transaction.

The broader MDK integration run retained 177 passing case rows and 14 failures while its historical current-module bindings were stale. Its nativeOne fixture created a Worker before module validation and leaked it on a validation exception before the response Promise existed. The first 191 case results and the exact own harness SIGTERM are preserved in mdk-integration-verified.log and mdk-integration-leaked-fixture-stop.json; this run is not claimed as normal cleanup. The test-only finally now terminates that fixture Worker and closes its port on every exit path. The repeated suite exits normally with code 1 and no signal: 177 pass, 14 fail, zero cancelled/skipped. All 14 failures retain the historical module-size checks: expected 2,154,315/2,156,930 bytes versus the new 2,157,350/2,159,965-byte modules. mdk-integration-summary.json lists every failure; mdk-integration-cleanup.log/json retains the raw errors and actual exit. No rejected module or original closure is admitted to turn a historical test green.

The unrestricted historical local-runtime suite still rejects its immutable original closure. That closure was already stale before this task (both D3DIM source hashes differ at the starting Git SHA); this work also changes host and observer bytes. The rejection is preserved, not amended into an original retry. The focused consumer regression tests the new contract and this closed guard.

First failures are retained: the build's three reviewed host census seals, initial proposal/fixture syntax and VM-realm setup errors, a fixture expecting three publications when ordinary composition produced five, the legacy fragment continuity rejection, and MDK's exact observer-source seal. Corrections changed source integration or synthetic fixture assumptions, not runtime caps, guards, original scheduling or acceptance. The first full legacy regression used its default small fixture-pin destination; subsequent evidence explicitly uses this assigned run. That deviation and the prior-byte unknown are retained.

final-cleanup.json independently rechecks exact run PIDs, normal exits for all passing regressions and the locked build, the repeated historical suite’s normal code-1 exit, the first leaked harness’s explicitly non-normal stop, watcher release, predecessor profiles/ports, both free locks and disk. publication.json records the final main/origin commit and strict-key pull/push result. No browser, server or sandbox was started by this task.

Actual setup input readiness, complete accepted pixels, personally reviewed fresh controls and gameplay remain unverified. Further original execution requires separate owner authorization and a new exact runtime binding. No box was assigned (ID/state/expiry: N/A); shutdown ownership remains with coordinator.