feat(application): read bounded source-local SDK clones

[pengfei-threemoonslab]

Application comparison can now read an OpenAI Agents SDK clone from a unique source-local SDK agent, retaining inherited tool, handoff and MCP lists and explicit overrides. Narrowing the clone produces rows for the clone while preserving the source agent.

Refs #911. Stacked on #1017 (codex/874-instance-tool-flow). This is a bounded source-local increment: imported/custom-instance/template receivers and full GRAAL acceptance remain open. Corpus scores, ledgers, goldens, baselines and released-build evaluation are deferred by the owner.

Boundaries

• At most eight clone hops; exact SDK constructor/method, namespace and retained-handle proofs remain required.
• A literal name override names the clone; other identities use stable source construction labels. Construction provenance is retained separately, so an unread replacement at the same site is uncertainty, never an established deletion.
• Literal function-tool name filters can narrow the list. Object display names supply no runtime filter proof; other conditions remain source text.
• Shallow shared lists retain their mutation limits across source/clone holders, including aliases, conditionals and second-level inheritance. New list literals, spreads and list copies do not become shared list containers.
• Optional agent_settings carries recognized source literals, not runtime authority. Tools remain bound under tool_choice="none". Conflicting or unread settings carry row-specific uncertainty.
• Default static scan does not expand clone observations. No baseline, waiver, policy or workflow was changed.

Authorship

• A coding agent wrote some or all of this change. Agent: Codex.
• A person read the whole diff before opening this pull request.

Validation

• 92 new clone tests passed, including real Git/CLI before/after comparisons, constructor and alias refusals, handoffs/MCP, identity movement, unread replacements, settings conflicts and shallow-copy mutations.

• Independent review passed 310 related tests and the final 214-test selection, plus explicit no-alias sharing and separate-new-list probes.

• Final stable constructor compatibility selection: 913 passed. After the final binding-scope absence correction, clone/source-identity/unobserved compatibility: 214 passed independently and locally.

• An earlier related selection had no assertion failures but six frozen-engine fixture errors because the implementation moved during that run. The affected constructor modules were rerun against unchanged source and all 913 cases passed.

• Ruff and diff whitespace checks passed.

• Original committed-source verifier/control at a63f0c817ef0: complete, release decision passed.

• Sequential restack at 8ed96b350e60: candidate-object refusals and candidate uncertainty annotations were retained alongside clone filtering and inherited handoffs. 94 candidate/clone tests passed before the unchanged scope-fix replay; extra independent combination review passed 383 tests and found no actionable P1/P2; candidate uncertainty, inherited conditions and refusal routes were preserved. Committed control is complete / passed; the first stacked CI exposed a stale clone-is-unread assertion; its independently reviewed correction is recorded below.

• CI compatibility correction at 3618a3077f31: the existing imported-tool fixture now requires all three supported clone/source changes, their exact source and empty uncertainty/gaps. Clone-method replacement and escaped-handle cases still require partial / not_established with side-specific refusal evidence.

• Complete imported-tool/clone compatibility selection: 289 passed; independent test-only follow-up: all 4 original parameter cases passed, including both refusal subcases. No production source was changed. Committed control is complete / passed; exact-head GitHub CI passed all 14 applicable checks, including six suite shards and aggregate coverage; the release-tag-only check was skipped.

• Six-increment combination at a4ea56e5987bf: 1,546 passed, 1 skipped. Final product source is identical after the test-only correction; independent combination review passed 383 tests with no actionable P1/P2.

Independent review and fixes

1. Fixed two P2s: same-name clones with different settings were merged/compared as certain; object display names were used as runtime filter names. Conflicting settings now taint each affected row; object-filter membership stays uncertain.
2. Fixed P2s in absence and setting evidence: replacing a literal/fallback clone with an unread factory looked like deletion; a model override could look like a known setting deletion. Structured construction and binding-target labels, including unpacked/loop/annotated assignments, imported objects and both syntactic/global/nonlocal binding scopes, now protect absence, and unread setting decisions carry uncertainty. Settings-only row explanations identify the actual changed declaration.
3. Fixed a P2 in shallow-copy sharing: adding a tool through a clone could leave the source falsely complete. The existing in-place mutation route now follows shared literal, explicit-field, alias, conditional and second-level list holders. Independent final review found no remaining actionable P1/P2.

CI follow-up: the reviewer independently reproduced the obsolete partial expectation, verified the supported source-local clone's three changes and refusal variants, then reviewed the exact test-only correction. No actionable P1/P2 remained. CI coverage correctly rejected the earlier incomplete suite; no gate, expectation for a refusal, or test selection was weakened.

The SDK clone contract is a shallow copy. No code was imported or executed to discover agent behavior.

Ordered merge preparation

This increment was replayed onto main 337fb5cd44eb, after #1017. New head 5daa9a061904 has exactly the same file content as reviewed head 3618a3077f31 (full-tree diff equality). Fresh committed control is complete / passed; exact-head CI for the replay passed all 14 applicable checks; release-tag-only check skipped. The preceding main commit passed all 13 CI jobs before this merge. Baseline implementation remains deferred.

Post-merge main verification

Squash-merged as 662236c775b29b4cec1d424e87818743a9fa0ee9. The main tree exactly matches the reviewed PR head. Post-merge targeted tests passed: 289 tests. Fresh committed verification on that main commit returned complete / passed. Main CI run completed successfully with all 13 jobs passed. Baseline, corpus, ledger, Q2 and released-build evaluation remain deferred.