VUH-2881: Struggle orb worker receipts

Source candidate; live acceptance remains with the lead. No game, desk, live, Steam/install or save action is part of this work. The review delta is based on main aa0d64c (VUH-2898's arrival wait).

What the historical run proves

Read-only input: Juno's .local/wt/vuh2466/build/scenarios/20261026-192703_net_prologue_struggle_orbs_1, main bc8d9ad. The host (6536) sent 242 birth rows; the friend (219608) received 242. The final host stats (:2418) show births=242 collects=0 ... offThread=41208; the friend (:2502) shows dropsSuppressed=38 picksSuppressed=0 offThread=44591. Its final display sample shows created=242 vanished=0. Both seats reached the match result.

The shipped Gate::Admit returned before every native read and booking callback whenever the detour TID differed from NoteGameThread's TID. Therefore all those foreign entries bypassed mirroring, including client-local suppression. Simply removing the comparison would call EnemySync's vector/book/outbox mutations and actor mappings from workers; those APIs are owned by the entity thread.

Limit: the old logs contain neither per-hook TIDs nor per-hook foreign counts. They do not prove which hook produced each foreign entry, nor establish a specific native worker TID. The native facts/pinned prologues identify the record argument as R8 for 40FF50, RDX for 410250, and the flight transition as update +0x10 plus collector +0x48; these facts alone cannot prove dynamic thread affinity. The evidence supports fixing the unsafe affinity assumption, not inventing a thread identity. Exact selected lines and original hashes are in tests/fixtures/struggle_orb_threads/sources.json; they remain failing controls.

Change

OrbTick publishes an immutable snapshot of the admitted session generation, delivery, epoch, native transition/load, manager, role and collector identities. A detour on any thread reads that snapshot. Native code runs once on its original caller. Workers never invoke EnemySync or resolve an actor through live containers. Unknown collectors are sent unresolved, so the friend fades the orb rather than guessing.

Pump on the noted entity thread drains copied receipts into the existing host book and emission outbox. Its book keys are monotonic incarnation IDs, not reusable record addresses. The mailbox holds at most 4096 receipts and 4096 record-address mappings; overflow is counted and fails acceptance. Scope/manager/role changes and room resets retire queued and in-flight receipts, including a manager returning at the same address. No lock is held over a native call or a callback. Worker client-local drops and picks use copied policy, rechecking the native manager and live policy before skipping.

Each actual detour (create, init, hit-drop, drop, pick-a, pick-b, fly) and NoteGameThread emits thread site=... tid=... gameTid=..., bounded to the first 16 TID changes per site. Periodic per-site totals retain entry, foreign and change counts. offThread now records observed foreign entries; it is not a failure or a worker relabelled as the entity thread. Native input/output LastError is preserved around tracing and receipt work.

Sorrel's ordering finding and the revised boundary

The first candidate's post-return FIFO claim was wrong. An allocator can publish record R, pause before submitting its birth, and be overtaken by another worker's collect of R. The owner then drops the unknown collect and eventually mirrors a permanent orb. Likewise an old pickup's delayed return can name a reused record. This is an admitted interleaving, not an observed failure in the historical run.

Read-only queries of the existing analyzed program establish a narrower native boundary (raw decompiles in .local/ghidra-orb-{review,lifetime,init,base}.log):

• 3DC950(manager, descriptor, point) calls 413F00 once and then sets scatter.
• 413F00 gets a pool record, calls 412D40(record, manager, descriptor, point, update), and only then sets the record's enabled byte +1A=1.
• 412D40 calls 4118E0, clears flags +28 and collector +48, and initializes the manager/descriptor and life. 4118E0 initially sets +1A=0. There is no native incarnation counter in those fields that a delayed outer return can use.
• Both Struggle pickups call 413650(record, actor, 0). It unconditionally writes that actor to +48, establishes flight life/update, and returns. Its arguments identify the collection without rereading a potentially reused record later.

The revision therefore hooks the pinned initializer and flight function too:

1. The initializer reserves an incarnation before native initialization, while the native record argument is known and before the caller enables it. The reservation belongs to the outer create call on that native stack; stack identity and exact arguments prevent a different fiber borrowing its TLS frame.
2. The outer allocator completes that reservation with copied scatter only if the incarnation is still current before and after the native read. A new initializer invalidates the old identity before writing any new record fields.
3. The flight detour captures the incarnation and copied collector identity at entry, reserves its collect, calls native, then completes that same ticket. A delayed return never consults the record again or adopts a newer identity.
4. The owner drains only a completed queue prefix. A collect cannot pass an unfinished birth, and a new birth cannot pass a reserved old collection. If collection/reuse happens before a birth becomes observable, the adapter elides it atomically, including its collect. elided reports those births. It neither reads flight velocity as birth scatter nor creates an orb already gone.

This preserves native call concurrency; it adds no lock or wait across a native call. It handles delayed observer returns rather than assuming KH2 never overlaps two workers on one pool address. All seven hooks apply together or stay inactive.

Controlled wrapper tests force two workers onto the same owned native record: allocation post-write pauses while another thread collects and the owner pumps; then native flight post-write pauses while another thread reinitializes that address. The first must elide without any wire birth; the second must send the old ID's collection and leave the new ID displayed. Sequential reuse remains a separate control. Scope retirement during the inner flight also must not book into its replacement.

The scenario now requires nonzero friend drop and pickup suppression, matching sets of sent/received (epoch, sequence, orb ID) collections, and a clone-owned collection with a successful matching friend application receipt. Zero suppression, mismatched IDs/epochs/sequences, an extra unmatched collect, missing/failed application and zero clone collections are negative controls. The receipt proves the native flight/fade call succeeded, not that a video shows its motion. Per-seat mapping of every site (explicitly naming unexercised sites), paired visual evidence and a deliberate clone touch opportunity remain live work.

Delta verification: after Juniper's batch ended and the lead cleared PC1, the branch rebased cleanly onto aa0d64c. git range-diff confirms the two previously reviewed commits retained their content. Fresh serial/background Release build passed; CTest 280/280 passed in 275.89 s, including 44/44 Python files, 46/46 actual-wrapper checks, 456/456 adapter/rule checks, and 10/10 scenario controls. The unchanged resync test passed in 15.83 s. Gate snapshots show no desk job before/after these jobs. The first focused wrapper run exposed a test-accounting error (the normal owner collector-snapshot callback was counted as an unexpected booking); the assertion now counts actual birth/collect callbacks, and the focused and full suites pass.

Fresh run.py --validate checks 118/122 fixtures, including the updated orb scenario. Its four failures are the unchanged, pre-existing unknown-do cases: net_player_progress_route, net_safe_field_recovery, net_twilight_thorn_legA, net_twilight_thorn_postfield. The delta receipt and native decompile excerpts are attached to VUH-2881; raw checks and gate snapshots are in this worktree's .local/review-* files. No native affinity, visual flight/fade or clone brush-past claim follows from these offline results.

Sorrel's application-retirement finding

The first review delta called OrbClientFollow() from the successful-flight callback inside ClientDisplay::Tick. That helper can observe a new runtime generation and retire the session, clearing the very display queue that Tick is iterating. Tick would then erase an invalid range. Comparing the epoch before that helper did not prevent retirement or suppress an earlier stale receipt from the same tick.

The callback now only invokes the native flight and copies successful requests into a stack array bounded by Tick's 64-call budget. After Tick returns, OrbTick refreshes authority, then compares the captured generation, epoch, delivery, transition/load and manager before emitting any application receipts or display summary. A retired scope discards the whole batch. The display's callback contract explicitly forbids reentrant mutation/reset.

The owned-memory NativeHitClaimTest runs the actual EnemySync callback. It creates two display orbs, queues both collects, and advances the live writer's generation during the second successful native-flight leaf. A reset guard rejects retirement before Tick consumes the queue, stopping a regression before the invalid erase. The reviewed code fails that guard and leaks the first application receipt; the fix safely retires the display and emits neither old receipt. Stable successful flights emit both receipts; failed flights emit none. The writer remains live and the seat remains a client throughout the race. This is a controlled native-leaf schedule, not a live game reproduction.

Fresh serial/background Release and CTest 280/280 passed in 274.41 s, including all 44 Python files. The focused production native harness passed 1,406 checks. Red/green logs and gate snapshots remain in .local/p1-*; the portable receipt is attached to VUH-2881. Scenario/runner files are unchanged, so the prior 118/122 validation (four known baseline failures) is reused.

Original candidate verification (historical)

• kh2coop_struggle_orbs_test: adapter controls and three real producer threads submitting 600 receipts concurrently with owner drain. Scope changes, unregistered/inactive context, client suppression, policy withdrawal during a read, stale return, native unwind, unknown collectors and bounded overflow.
• kh2coop_struggle_orbs_hook_test: actual private production detours, native reads, TID logger and Pump over owned memory and synthetic native functions. Both pickup argument routes pass births/collections through the real codec, ledger and client display. The displayed orb flies to its copied collector; same-frame pool reuse preserves the newer orb. A blocked native call does not block owner policy retirement, and its late return is refused. No hooks install.
• test_struggle_orb_threads.py: executes the committed scenario predicates in the production runner Context. Historical excerpts fail; synthetic positives allow foreign entries only with thread evidence, drain, sent/received collects, friend vanish and no overflow. Missing evidence/collect/receive/vanish or either seat's overflow fails. Synthetic positives are not a replay of unlogged pickups.

Source 2a2a9d0 on main c330cfd: full serial Release build passed. The full CTest run passed 278/279 in 268.48 seconds, including 43/43 Python files, 35/35 actual orb-wrapper checks, 447/447 adapter/rule checks and all five new scenario-predicate tests. The unchanged resync target failed one combined quarantine assertion in 19.77 seconds, then again in 19.60 seconds.

Its original machine-code section matches the passing main binary, with matching compiler, build graph and include sets. After restoring temporary diagnostics, the original binary passed 308/308 checks, 3/3 idle runs, in 15.97, 15.45, 15.51 seconds, with the same CWD and no desk job before/after each. The main baseline passed in 15.55 seconds. The timing-sensitive failure is separately tracked as VUH-2894; both red logs remain preserved. The original failures preceded Juniper's base job, so attributing them specifically to that live run would be unsupported.

The scenario passed run.py --validate before rebase; the fixture and validator blobs are unchanged, so that passing evidence is reused. A fresh recheck was deferred when the next live job began. All builds/tests used below-normal, serial execution; every build/test after the live handoff checked gate/jobs first. Sanitizers are not claimed; this uses the normal MSVC path.

Offline verification receipt is attached to VUH-2881 and was downloaded back byte-for-byte. It pins source, staged DLL, historical provenance, build/test logs, synthetic TIDs and idle trials. Raw logs remain under this worktree's .local/; the evidence-store CLI is unavailable in this seat. The attached receipt and issue are the shared record.

Remaining qualification

Independent inject review, then the lead's net_prologue_struggle_orbs rerun. The new per-site TID lines must name actual native hook threads, nonzero host collects must reach the friend as flights/fades, and the clone's brush-past question must be observed. Offline synthetic threads prove that the wrappers handle foreign callers, not which KH2 scheduler thread will enter them live. Saves must remain unchanged; no live acceptance or Thorn victory is claimed.