Conan burst-safe writer observation, 2026-11-06

Bounded transport now survives the generated and original loading bursts without losing admitted calls or returns. The newly measured blocker is earlier global detail/admission exhaustion by repeated count 103 constructor epochs, rather than count 2 owners or transport backlog. No emulator defect or player-controlled gameplay is established. Production code and original guest bytes are unchanged.

Evidence: scratch/runs/20261106-conan-burst-safe-writers/ (ignored). The preceding investigation lacked the earlier history of its first-zero owner. Its zero time versus channel starts45.900, later weight1, and11,196 native-matching stores on other owners did not establish an emulator defect. Old pointers are not identities in this run.

Coverage changes

Run-owned compiler observers leave production source and original guest bytes unchanged. Admission requires actual logical count 73..65535. At constructor 0x431950 entry the owner is uninitialized, so admission reads its model argument, matching the constructor's call to 0x412c80 (MOVZX EAX,WORD PTR[ECX]). Later admissions use owner+12. Repeated count 2/count 72 owners consume no lifetime slots.

The auxiliary-copy predicate normalizes nesting depth before boolean AND, covering both even and odd nesting. Index72 copies require logical admission and contain only its own 16-byte entry. Copied metadata includes admission count and owner/table backing addresses for actual store-destination checks.

Each admitted entry reserves its return slot independently of detail capacity. Pending and in-flight sink work remain charged until persistence completes. Saturation refuses admissions/details explicitly and preserves admitted ends, including at the inherited packet quota. Histories with missing or mismatched endpoint/owner/parent/thread/PC/SP ends, detail loss, or persistence failure are rejected. Scope remains bounded: decoder eviction, passive watch replacement, unobserved writers and owner reuse can still prevent a complete causal history.

Executed controls

The locked current build passed. Standard and compatibility baselines match shipped bytes; diagnostic modules preserve import/export contracts.

Validation Measured result
Public COM, actual owning CPU/render Workers 20 baseline/diagnostic cells pass
Generated loading burst, both modules 15,340 persisted rows per diagnostic module; 2,051 entries and 2,051 matching returns; no pairing errors
Logical admission and nested auxiliary copies 6,000 count 2/count 72 sequences excluded; count 73/count 103 admitted; depths 2/3 copied
Delayed drain and bounded saturation Reserved ends survive; new admission refused; incomplete/bad-PC-SP/sink-failure histories rejected
Genuine MEM_RESERVE/SEH within admitted lifetime 16 cells pass across both modules, eager/lazy and baseline/diagnostic
Independent native x87 24 multiply/FCOMP and12 envelope controls pass

Architectural registers/lazy flags, x87 payload/environment, child-thread and fault state remain equivalent. The strict singular-normal rendering error is preserved. Burst detail8192 and invocation2048 limits produce explicit rejection; the compatibility summary's truncation list is cumulative, not a lost-event count.

The first generated fixture attempt overlapped its enlarged prelude with the old child trampoline. This was a generated-input defect, not an emulator defect. The failed attempt and original pins remain under initial-*; moving the trampoline to free 0x42ab00 and asserting prelude geometry fixed it. Corrected controls were rerun before the original.

Runtime and recovery

The exact AoW0410 → D3D0430 → Aranna0500 final releases were checked in order. Independent cleanup verified247 absent PIDs,17 refused ports,8 absent profiles and4 free locks. The first collector enumerated4 profiles; the expanded check covered all8 before any shared edit/test/build/runtime and retained the original immutable05:57:09–06:42:09UTC window, including cleanup.

The build/module controls used Git 36b8c13a6927bb2752a16677921bf376716bb792. The original source pin is 41eeb8ebfce265ab40f8e679d67ed0ee667fe7e8, a later documentation-only commit; canonical source/helper/module closure was rechecked before launch. Recovery material is exact Git SHA, relevant.binarydiff, new-source patches and helper-new-file-patches/. Media is referenced by existing path and SHA256 in media-references.json; no source archives or media copies.

Actual page, owning CPU Worker and rendering Worker module identities all match the diagnostic bytes; 301 source-file checks and 18 served-JS identities pass. Post-run closure/helper/module checks and all 37 existing media hashes pass. No boxes assigned or used; no lifecycle actions.

Single original: measured boundary

One session ran06:02:50–06:19:19UTC and closed normally, including Chrome exit0. Personally inspected ordinary input advanced New Game to the Warrior difficulty menu and then Kordava loading. The final inspected client was black, with no retained rendering failure. Movement plus another visible response was not obtained. The original was stopped after the coverage boundary was measured; it was not restarted or continued in an attempt to recover already lost history.

epoch-audit.json verifies3,497 consecutive durable segments,32,014 rows, consecutive transport ordinals,2,048 exact admitted entry/return associations, zero open admissions and zero mismatched returns. The last capture was drained, with zero transport rejection, no sink failure and high-water638 rows; four subsequent observation rows were persisted by the final flush. Transport capacity is16,384 detail rows plus65,536 boundary rows, with reserved ends charged while pending or in flight. These are diagnostic bounds, not increased guest memory.

Every admitted original call had logical count 103. All used reused owner address 0x08830d80;51 constructor epochs were retained. Function admissions were51 constructors,51 setups,948 time updates,948 evaluators and50 0x430fa0 calls. Address reuse is explicitly separated in the offline epoch audit. Nineteen retained epochs have matching observed calls and no explicit quota loss; this does not establish destruction, all possible writers, or any first-draw link.

First retained packet reporting a bound Exact observation
Detail loss, serial13,888 / transport ordinal16,000 0x4310a0, invocation808,809 entries,808 ends,8,192 details; count 103 owner 0x08830d80
Admission loss, serial22,627 / transport ordinal24,739 2,048 entries and2,048 ends already completed; explicit overflow flag

These are first retained loss reports, not an asserted instruction for a missing event. The earliest full epoch shows actual ordering:

Retained serial Actual observation on that epoch
1→48 Constructor 0x431950, matching return PC+ESP
4 0x43195d writes owner+4 to zero
34 0x4319fd fills the table, including index 72, with ffffffff
36 0x431a13 clears the channel span including both+38/+68 weights
49→99 Setup 0x4314f0, matching return
95,97 Two actual 0x4316da word stores in index 72's16-byte entry
100→130 Time update 0x4310a0; nested evaluator111→118 returns at 0x4311c1
165 and later Actual 0x4310b8 stores change owner+4

Store claims require adjacent pre/post packets and actual copied backing-address overlap. The raw packets retain operands and values. Native arithmetic was validated by the generated controls; no new original arithmetic or scheduling mismatch is claimed. None of these temporary epochs is tied to a first index 72 palette or scene draw. No index 72 palette observation or first-zero target was captured before shutdown. The prior target's ownerTime0/channel-start45.900 question therefore remains unanswered. Later decoder eviction is reported (103,275 at the final capture), further limiting instruction attribution. The separate world/draw timeline had already reached its inherited bounds (2,048 world rows and64 draw rows) during menus; it cannot establish later scene-draw ordering either.

The next useful investigation must preserve relevant constructor epochs through the first draw within the existing bounds, accounting for these repeated count 103 temporary epochs. Increasing memory, replaying this exhausted observer, or repairing guest animation values would not answer the missing causal link. There is no demonstrated x86/Win32 divergence to regress or repair yet.

Inspected media and closure

Paths below are relative to the ignored evidence directory. Every inspected PNG and its SHA256 is listed in inspected-captures.json; no media was copied.

PNG SHA256
new-game-held.png — Warrior difficulty 351ccf443e882096e8b1351584b27d5034f0dc8c75f41f91362aaf88c962ec9f
difficulty-warrior.png — Kordava loading 6f1feb68ed236ee6b583758d9a91f17e0d8cf2c2fd4a3343d138317fb87c2fc3
loading-11.png — black client, no gameplay 36cc8fe8c45072720d525204ac11524d172ae595f4cb12ca9b344523af9a3667

The run-only analysis addition has its own analysis/epochs.py.patch; runtime helper pins remain unchanged. runtime-evidence-verification.json, postoriginal-closure.json, owned-runtime-cleanup.json and the final artifact manifest provide identity and cleanup checks. result.json is written last, after publication and reopening/hash-checking the retained artifacts. Only this note is committed; the commit names this run and the executed tests.