JavaScript assignment influence verification — 2026-10-24¶

PR #6206 answers which retained inputs can influence a reassignment. It links reads to binding/property writes, retains both compound operands, and keeps conditional inputs and resulting values uncertain. It does not evaluate arithmetic or prove execution order.

Source and ownership¶

The comparison baseline is b25895f9522a3432f0c2f9dfd2fc3e819510298f, the original assignment feature integrated with main a062529b. The corrected runtime is 3224acae383501c5559c43ef521220406f3fee5d. This compares two implementations that both answer the ordinary fixtures; main without the feature cannot answer all reassignment questions and is not a completed equivalent baseline.

The parser now retains nullable computed-key source ranges on callable facts. The shared graph owner lookup excludes those ranges from the method while retaining the method's full source identity. Domain traversal and graph lookup reuse the same computed-method predicate. Parameter defaults and method bodies keep method ownership; callbacks nested inside a key keep their own ownership. The strict worker schema carries the parsed field across processes. Graph consumers do not reparse the source or special-case assignment owners.

Assignment collection uses the shared cooperative traversal (a checkpoint every 1,024 visited nodes). Per-projection read/write range indexes reuse retained nodes and have the immutable file projection's lifetime; no cache was added.

Executed correctness checks¶

• Three new object/class/enclosing-function computed-key cases fail before the ownership fix (11 pass / 3 fail) and pass afterward. Parameter defaults and nested deferred callbacks are covered. The final assignment file has 15 passing cases, including an assignment genuinely owned by a callback inside a computed key.
• Removing the assignment owner filter temporarily makes both deferred-callback negatives fail (12 pass / 2 fail in that control). Candidate traversal is explicitly enabled and the graph is checked for erroneous assignment edges. The production file was restored before further verification.
• Full npm run check passed after integrating main 1a48a820 at 0fbd13b77b15c6906138eaac232456d3a2141ac5. The initial 14-file run had 295 pass / 1 fail: graph construction/scope/destructuring, semantic analysis, worker schema/transfer/budgets, 24 cancellation cases, and the production CLI fixture workflow.
• That remaining projects literal seeds and static object flow assertion expects 3 matches but receives 4 on both this feature and exact main 1a48a820208880c0cc383d196c7ed96d35479412; the diagnostics match after checkout path normalization. The historical failed assertion is not counted as passing in that run.
• The CLI test executes analyze plus default and candidate-expanded traces for declaration, variable, property, compound and computed-key fixtures. It verifies the intended write and reachable return through the public commands. Full documentation and test-discovery checks passed. No target code is executed. MCP runtime delivery and other platforms were not newly verified for this change.

Later integration and verification¶

The integration with main 1c07bee5b6262a537758e8bfa235e90d99e35d71 produces the same tree as actual CI merge 49527572db7de2dfb260344f8c69c74f7b466423. Full npm run check and all 296 selected feature/worker/CLI/cancellation checks pass on that tree. Upstream changes repair the earlier literal-seed assertion recorded above.

A cancelled shard also exposed a stale introductory-description assertion in analysisViewMcp.test.ts. It prevented the oversized native discovery workflow from running. Removing that unrelated catalog/prose precheck preserves the metadata filters, 6 MiB retained-dossier discovery, paging and advertised-schema checks; all 22 tests in the file pass. The unchanged original case fails before this cleanup. Full static checks pass with the revised test.

Separately, the CI source lane's 14 failures reproduce on both exact main and its actual merge (each 167 pass / 14 fail across 181 checks). Twelve shared-child failures have matching diagnostics; two 2-second timing assertions fail on main as well. Failure multisets match after normalizing checkout paths and the elapsed number only; timing measurements themselves are retained, not claimed equal. These independent main failures are recorded rather than reported green. Performance costs of this later integration were not remeasured.

The isolation audit also found that the headroom mock test added in #5660 was incorrectly placed in the shared services project. It and the existing MCP cancellation mock now live in tests/process-global/, as the established test policy requires. All original budget, cancellation, retained-fact and subsequent request assertions remain. The cancellation fixture holds the selected progress boundary until an ordered client ping acknowledges delivery of its cancellation; cleanup releases the gate even on failure. The fixture checks that this gate was actually used. All 18 headroom, SDK-cancellation, progress and project-policy checks pass. Bypassing the fixture's producer gate makes both cancellation cases fail; restoring it preserves the passing results. An earlier two-file preload control passed and does not prove import order caused the CI failure. No CI speedup or new real-provider verification is claimed for these test moves.

Completed task comparison¶

Each task starts two fresh production CLI processes: analyze the same source folder, save its complete Evidence, then trace forward from literal 0.001 with candidate edges enabled. The source oracle expects 1 or 32 reachable returns. The 32-function fixture contains 96 syntactic assignments and retains 128 assignment relations, 386 nodes, 481 relations and 192 unknowns; its ambiguity remains visible. All six semantic answers per fixture compare equal, including identities, provenance contexts, unknowns and coverage. Only the retained parent Evidence identity/link fields are excluded from equality.

Three pairs per fixture alternate before/after order. These are fresh process runs and repeated filesystem-cache runs, not warm analysis-cache hits. OS file caches were not flushed; the host was shared. Medians:

Output token estimates are effectively unchanged. They use pinned tokenx 1.3.0 on both complete CLI responses, not model usage. The saved query file repeats complete Evidence on disk: 13,440 / 402,585 additional estimated tokens for the respective fixtures, unchanged in both runtimes. These are separate file payload costs, not agent context usage. Known-command runs do not request a help/schema/catalog. Invocation text, prompts, model framing and rendering are unmeasured. Estimation runs after the timed task.

No speedup is claimed: small elapsed ranges are 4.857–6.328 s before and 4.756–6.426 s after; 32-function ranges are 5.719–9.262 s before and 5.871–10.679 s after. The latter CPU median increases by about 0.30 s (4.0%) and peak-RSS median by about 4.5 MiB (0.8%). This is a correctness and cooperative-cancellation change, with the measured cost retained explicitly. These small source fixtures do not establish multi-megabyte application performance. Later main integration costs were not remeasured.

Reproduction and accounting¶

Linux, Node 24.22.1 / npm 11.16.0. Node processes use CPU affinity 0–1 and 2 GiB heaps. The observer samples recursive process-family RSS every 50 ms, with observed 6 GiB RSS, 180-second CPU and 300-second wall stop thresholds. All 12 tasks exit successfully without a resource stop and preserve source bytes. Timings include both CLI invocations, JSON parsing, saved Evidence/query writes and the source oracle. CPU is waited-descendant accounting. RSS sums shared pages and may miss short-lived or unobserved descendant transitions; it is not a kernel-enforced memory ceiling. Fixture hashes:

• fixtures/small/main.js: 4e69e2708b212d1659eb0372e4fb5942e00906b2e9b44b3fa3e1e54eddb13ecb; 94 source bytes.
• fixtures/large/main.js: faea1107b93c09fd2594049ddac60425d16ad7c2989a912377332fc84373dc68; 4,875 source bytes.

Small source: const accel = 0.001; let speed = 6; export function update() { speed = accel; return speed; } followed by one newline. The other source starts with const accel = 0.001; and a newline, then joins the following line for i = 0..31 with newlines, including a final newline:

export function update{i}() { let speed = 6; const state = { speed: 6 }; speed = accel; state.speed = speed; state.speed += accel; return state.speed; }

For each pinned checkout, build with the locked toolchain, then invoke node scripts/rea.mjs analyze-javascript-application <same-folder> --format json --full-output. Save its envelope's data as the application in a query with seed: {kind: "literal", value: 0.001}, direction: "forward-influence", include_ambiguous_dynamic_edges: true; run node scripts/rea.mjs trace-javascript-semantics <query.json> --format json --full-output. The task harness SHA-256 is 93a2b5908013bdaaa39cc1492a61fc159fc71874fc7dad51a74a57f9f5d6f346. Raw command receipts, stdout/stderr, equality checks and all 12 measurements are retained in the local maintenance evidence directory; generated runtime, large response files and local planning artifacts are not committed.

Current-main integration follow-up¶

Integration 8f7b403a0b1349700a819eda373d3dcac0e6133f includes main 85e38ac199dff15e8c43343d9170db7e0e46d284. Main independently corrected the old discovery assertion and added mock-isolation path overrides. The merge retains main's SDK JSON Schema validator and tool-name discovery check, updates the isolation policy's two path assertions to the moved fixtures, and removes now-obsolete project includes/excludes. No test is assigned twice.

The full static gate passed (63.836 s elapsed, 74.452 s waited-child CPU). The expanded selection passed 526 checks, including all assignment/ownership, worker, cancellation, 22 MCP view, headroom and 14 policy cases. The formerly failing source/CLI baseline cases pass on this newer main integration. One public assignment CLI case failed at process exit while this operator had incorrectly allowed the static gate's runtime compilation and built tests to overlap: its start was 34.664 s before gate completion. Its output did not retain the subprocess diagnostic, so the specific subprocess failure is unknown; this is not evidence of a main regression. After build completion, the complete five-fixture public assignment CLI case passed in 34.220 s (41.792 s CPU). The fixture now includes subprocess stderr and the fixture name in all exit assertions. The failed attempt is retained and is not counted as a passing run.

Earlier head 4abe5509 CI run 40048477593 finished with 11 successful, 4 failed, 19 skipped and 3 cancelled jobs. All 7 failed/cancelled logs, including aggregation gates, were inspected: 62 visible assertion markers were covered by exact-main controls; no unknown markers or failures of the repaired fixtures remained. The exact CI merge 3c8ed5cc and its main parent d81d6b67 reproduced both CLI failures with identical failure-message multisets. The new public assignment CLI passed on that CI merge. These controls apply to that executed revision; current head CI must be assessed separately, especially because main has since repaired several baseline cases. Cancelled jobs do not establish unexecuted coverage.