Refine Java overload ambiguity with resolver-owned argument types

[aksOps]

Summary

Stream C is the Java consumer of C0 (#60) and C1 (#61). C0 added optional argument types to Reference/RefKey; C1 made use of that key dimension resolver-owned and default-off. This PR populates a Java argument vector only when every argument is inside the approved file-local Stage-1 surface; unsupported or unknown shapes keep None. It refines a key only after the authoritative legacy pass returns AmbiguousOverload.

The typed retry preserves strict, loose, then varargs phase order; exact-array fixed arity; zero-tail varargs; integer radix, suffix, and range; unary numeric promotion; canonical java.lang spellings; and the stable legacy-first candidate union. Unsupported or unknown vectors remain honestly ambiguous. Java graph revision becomes 1; every other language remains at 0.

The planning card's pre-measurement AmbiguousOverload <= 2,800 estimate is superseded by the owner ruling and the measured residual boundary: 3,152 commons-lang AO occurrences have no argument vector inside Stage 1. Their calls, nulls, lambdas, member expressions, array access, and general operators remain assigned to Streams H/I rather than being guessed here.

The final-heavy array regression is also repaired: declared array receivers stop at NeedsTypeInference, while true local and parameter receivers remain LocalBinding. Commons-lang's 71 rows/75 occurrences and gson's 12 rows/12 occurrences retain their full C0 keys.

This diff changes Java-track implementation, Java tests, documentation, and attributed Java baselines/pins only. It makes no core/schema/model/store/pipeline, unresolved-reason taxonomy, dependency, or network-code change.

Attributed corpus results

Corpus	Resolved	Unresolved	External	Local binding	Rate
commons-lang	34,217 -> 38,222	16,279 -> 12,274	63,385 -> 63,385	15,162 -> 15,162	67.8% -> 75.7%
gson	12,885 -> 13,306	6,105 -> 5,684	16,737 -> 16,737	6,706 -> 6,706	67.9% -> 70.1%

Corpus	AmbiguousOverload	NeedsExpressionType	NeedsTypeInference	NoMatchingDefinition	UnindexedSupertype
commons-lang	9,218 -> 5,213	6,566 -> 6,566	342 -> 342	123 -> 123	30 -> 30
gson	1,282 -> 861	4,713 -> 4,713	72 -> 72	37 -> 37	1 -> 1

Occurrence totals are conserved: commons-lang 129,043, gson 42,433, and Java probes 22. The whole-row join reports zero changed non-AO rows and zero changed non-AO occurrences.

The movement units remain separate:

Corpus	AO -> Resolved rows / occurrences	Retained refined-AO rows / occurrences	AO -> AO common-full-key payload reclassification rows	New unique graph-edge identities	Removed graph-edge identities	Existing resolved targets re-aimed	Existing pin rows re-keyed	Changed non-AO rows / occurrences
commons-lang	1,768 / 4,005	571 / 2,061	317	1,702	0	0	0	0 / 0
gson	286 / 421	302 / 321	21	274	0	0	0	0 / 0
Java probes	0 / 0	0 / 0	0	0	0	0	0	0 / 0

Corpus	DefKind::Alias census	Existing pins held	New resolved pin rows	Vanished	Moved	Re-keyed	Final pin rows
commons-lang	390 -> 9,617	16,111	1,768	0	0	0	17,879
gson	30 -> 3,083	9,075	286	0	0	0	9,361
Java probes	0 -> 9	n/a	0	0	0	0	n/a

Only the Alias census changes; every other definition kind is unchanged. The new entries are synthetic runtime forwarding identities for callable signatures, not a new definition category. The reporting audit found no alias/node cardinality entering scan reports, JSON/MCP output, baseline counts, or rate denominators.

Rebase provenance

The writes used the release binary built from signed commit babb1f507b2a1ad25b7318595d1d31188ffcd821, SHA-256 f5c39ebbc25edcd904b3d9369270aac7835847ae6d9b81e2ab97b7adb69ae87b.

The exact commands were:

/tmp/arthron-0.1.0-orchestrator/stream-c/final-heavy2-rerun/branch-target/release/arthron gate corpus/java/commons-lang --language java --baseline baselines/java-commons-lang.toml --rebase --commit 598dfc1
/tmp/arthron-0.1.0-orchestrator/stream-c/final-heavy2-rerun/branch-target/release/arthron gate corpus/java/gson --language java --baseline baselines/java-gson.toml --rebase --commit 3ff35d6
/tmp/arthron-0.1.0-orchestrator/stream-c/final-heavy2-rerun/branch-target/release/arthron pin corpus/java/commons-lang --pins pins/java-commons-lang.pins --write --commit 4ed8b2c1c974afafc255df3390543931cea3b938
/tmp/arthron-0.1.0-orchestrator/stream-c/final-heavy2-rerun/branch-target/release/arthron pin corpus/java/gson --pins pins/java-gson.pins --write --commit 4ed8b2c1c974afafc255df3390543931cea3b938

The --commit values above are the exact committed provenance strings; baseline and pin formats print but do not verify them.

Recorded verification

• cargo fmt --check: rc 0.
• cargo check --tests: rc 0.
• cargo test --test java_resolve: 47 passed.
• Direct array/varargs suffix guard: 1 passed, 874 filtered out.
• cargo test --release --test java_corpus: 6 passed.
• Commons-lang and gson edge-pin tests: 1 passed each.
• README/baseline parity test: 1 passed.
• Literal committed corpus gates: 29/29 returned rc 0.
• Java implementation, array repair, alias reporting, and attributed-rebase reviews: CLEAR.
• git diff --check origin/main...HEAD: rc 0.

Final gate

The earlier full-suite run stopped on the subsequently withdrawn, unmeasured AmbiguousOverload <= 2,800 estimate; it is not reported as a pass. The final gate used clean commit b2d82b7f3dda817310be2344b4b4345286593f11 and finished PASS_WITH_DEFERRED_WARM_TARGET:

• cargo test --release --all-features: rc 0.
• cargo clippy --all-targets --all-features -- -D warnings: rc 0.
• Mutation suite: 27/27 mapped behavioral mutations killed, zero survivors in the reported set.
• Warm-upgrade and twice-clean determinism: complete stores equal across files, nodes, collision dispositions, rows, edges, candidates, and supers for commons-lang, gson, and the Java probes.
• Source hand-checks: 15/15 matched the recorded argument vector and full selected signature; 8 commons-lang and 7 gson sites.
• Product-network observation: strace -f -e trace=network returned rc 0 and recorded no network syscall; the changed production-file manifest contains no network-code path.

Performance used /usr/bin/time -v with taskset -c 0,1. The small-repository budgets are cold <=5 seconds, warm <=0.3 seconds, and RSS <=128 MiB; every Flask and Express run passed:

Side/corpus	Cold wall / RSS	Warm wall / RSS	Result
origin / Python Flask	0.71 s / 21,772 kB	0.10 s / 10,896 kB	PASS
branch / Python Flask	0.74 s / 21,644 kB	0.10 s / 10,636 kB	PASS
origin / JavaScript Express	0.90 s / 15,232 kB	0.10 s / 10,252 kB	PASS
branch / JavaScript Express	0.89 s / 15,104 kB	0.10 s / 10,752 kB	PASS

At 5M+ lines, cold peak RSS must remain below the 524,288 kB hard ceiling; 450,000 kB is the separate program-margin target. Cold <=60 seconds/MLOC and warm no-change <=1 second are timing targets, not Stream C gates.

Run	Corpus/size	Wall time	Normalized time	Maximum RSS	Result
Cold	Kubernetes / 5,353,211 lines	108.80 s	20.32 s/MLOC	294,672 kB	PASS: hard ceiling and margin
Warm no-change	same	12.29 s	n/a	125,772 kB	DEFERRED_TARGET_MISS: Stream B owns the <=1 s target