Q3 debugger source inventory, 2026-11-15

The warranted source-inventory correction was published as ac8fb9b43 on main after actual Chrome, current native qualification and the full flock build passed. The single authorized changed-route original demonstrated verified source reuse but closed at a required snapshot evaluation deadline before consent. No further original is authorized on this route. Heavy work followed personally verified TIM smart handles and LS32 source/build/runtime releases.

Evidence: scratch/runs/20261115-quake3-source-inventory/.

Demonstrated repeated source and causal limits

The previous original's actual original/operations.jsonl contains 412 distinct script IDs whose source responses have the exact 47-byte controller snapshot expression SHA-256 000ba5e8856b8d2a95512815f84b8ee419cdcb73dbe015ee87dc832cc883f31c. All belong to the same render Worker session and target, with 413 matching controller evaluator requests. There are 411 preterminal source responses and one passive late source response, for script 423/request 1307 at 3906.575 ms. That source request triggered the fixed three-second UNKNOWN. The pending snapshot evaluation 1309 arrived at 6334.850 ms. No postterminal request occurred. The previous route is closed. Queue, inspector dispatch, native occupancy and source-body timing remain UNKNOWN; there is no evidence of an engine bug.

Historical logging retained context IDs and script IDs but omitted the actual scriptParsed.hash. We cannot retroactively claim those events qualify for reuse. Actual source-byte repetition is demonstrated independently of that gap. Removing repeated source requests cannot guarantee that the still-required snapshot evaluations finish within their unchanged three-second deadline.

Identity and necessity

The primary CDP Debugger schema defines scriptParsed.hash as a SHA-256 content hash and specifies separate source and WASM bytecode return fields for getScriptSource. The same schema describes numeric context ID reuse across processes and the stronger unique context ID. Debugger enable also reports existing uncollected scripts, so an anonymous event alone does not establish evaluator provenance.

V8's primary inspector implementation uses GetSha256Hash when available and has a UTF-16 fallback. Its primary script hash implementation hashes UTF-8 text. Fresh Chrome actually reports the UTF-8 hash for the exact 47-byte snapshot expression and the Unicode control. The helper validates the public hash against actual fetched source using UTF-8 or the UTF-16LE fallback, records the validated convention separately, and always retains the actual UTF-8 source-byte hash. Unrecognized conventions fall back to fetching. The initial UTF-16-only candidate safely fetched every script and failed its reuse control; that rejected qualification is retained. No original ran on it.

Only known controller expressions sent by the current connection can enter the reuse cache, after a first actual fetch exactly matches the expression and validates its metadata hash. Reuse preserves each new script ID, numeric and unique context IDs, session and target, and identifies the first fetched script. This establishes source equivalence, not a newly fetched bytecode proof or a claim that every anonymous script originated with the controller.

Production URLs, modules, live edits, WASM, unknown languages, distinct or unknown anonymous source, malformed/missing hash and missing unique contexts still fetch. No source is inferred from URL or length alone. Session/target or context replacement invalidates reuse; failed, stale and postterminal replies cannot seed the cache. Cache and known-expression memory are each limited to 64 entries/256 KiB with a 60-second lifetime, connection-local only, with limits shared across Worker sessions.

Current validation

Fifteen independent fake-protocol tests pass, including 412 repeated exact snapshot expressions with one fetch, equal-length distinct anonymous sources, actual module-byte hash controls, malformed metadata, context/target/session replacement, bounded lifetime/capacity, fetch failure and sticky terminal/late replies. The final actual Chrome control passes all 16 tests. On one Worker/context, six unchanged snapshot evaluations caused six requests under the old per-event policy and one under the candidate; each new script ID retains attribution. Unknown nested anonymous scripts still fetch, two equal-size distinct WASM modules retain actual byte hashes, and Worker replacement/two real isolated contexts fetch independently. Actual waits/EOF and zero postterminal requests are verified. An initial successful control and rejected fixture iterations remain explicit; a delayed Worker bootstrap exposed and corrected a baseline fixture assumption rather than being labeled a controller script.

The current native qualification passed queued key and Mouse1 input, independent UP, immutable capture, complete visibility transitions and sticky UNKNOWN with zero postterminal requests. Its direct native/watchdog/supervisor-child and Chrome actual waits/EOF completed; all 20 recorded PIDs were absent. Eight visibility tests, the browser helper tests and flock /tmp/wine-build.lock bash tools/build.sh passed. The correction was committed and immediately strict-pulled/pushed before original admission. The complete native fixture consent image was personally reviewed before its finite host-only click.

Single original: exact partial result

The fresh original used the unchanged registered arguments, real ticks and cooperative 10000, with current source/module/media/import/TLB/font/resolved-PATH and native closure pins. The independent 1820-second watchdog was armed around the 1800-second native deadline; the 180-second Quit reserve remained unchanged. No additional original was launched.

Actual original inventory records 397 scripts: 33 actual source/bytecode fetches and 364 verified reuses. The exact 47-byte snapshot expression was fetched once and reused for 364 later script IDs in the same live context, preserving each script's attribution. Both actual WASM scripts were fetched: 2,209,708 bytes, SHA-256 47fbf18303efc59192ed3b5dbfcb48c0f27e5d70ee7ccc969fc6321f7ff72efa. Public WASM metadata was not substituted for those bytes.

The required render Worker Runtime.evaluate request 808 reached the fixed three-second UNKNOWN. Its already-arriving response was logged passively at 3605.703 ms; it did not reopen the route. The last preterminal public GL snapshot showed one nonsoftware WebGL2 context, zero draws and zero presents. No consent image was produced, no host click or game input occurred, and menu, arena, frag and native Quit are all unproved. This is a source-inspection correction with an exact partial gameplay outcome, not a demonstrated gameplay fix. The residual queue/body/inspector/native-occupancy cause remains UNKNOWN.

The route remained permanently closed with zero postterminal requests. Host teardown completed and direct native (exit 1), watchdog (exit 0), supervisor child (exit 0) and Chrome actual waits/EOF were verified; all 13 recorded original PIDs were absent, with no pin drift or HTTP byte mismatches and no remaining profile. Across this task, 54 named actual wait receipts and 109 recorded PIDs were checked; historical/ancillary detached-outer gaps remain explicitly UNKNOWN. Runtime and build locks were released before this report update. Evidence is indexed in result.json, including selected screenshots and original-terminal-analysis.json. No boxes, workers or deployment were created. Next work would require a new causal investigation and authorization; this assignment stops without retry.