Close the P0 completeness gaps: oracles, network, auth, navigation, and a re-runnable suite

[wlu03]

Implements the twelve gaps from the repo audit. Test suite goes 360 → 589 passing, ruff clean, tool reference regenerated.

What was actually broken

Three of these were features that already existed and silently did nothing:

• ReproSpec.oracle was never populated. build_repro_spec was called in three places, always without an oracle, and report_issue had no way to accept one — so _eval_oracle always returned None and every re-verification fell back to digit-normalized summary matching. report_issue now takes assertions, and auto-filed findings inherit the last set evaluated by check_assertions.
• audit_dom no-opped on the entire local path. LocalElectronAdapter never overrode it, so on the default no-API-key config the deterministic-evidence tool returned {} and read as a clean pass. Now implemented over CDP, with axe-core 4.10.2 vendored and injected as source text so an app's script-src CSP can't silently kill it. axe_ran / axe_error are forwarded to the MCP boundary — an empty violations list can no longer be mistaken for a pass.
• The fix loop was hidden. update_finding_status and verify_fix were in ADVANCED_TOOLS, so under the default core profile an agent could file a finding but never close it.

New capability

• Network — CDP Network domain captured as a bounded, failure-preferring channel; failed requests and 5xx file as findings; AssertionKind.NETWORK is a real check instead of a hardcoded inconclusive.
• Auth — cookies and web storage captured and replayed (capture_state / seed_state), so a run starts logged in instead of burning its iteration budget on a login form.
• Navigation — navigate / back / forward / reload, plus hover, right_click, drag destinations, and scroll direction. drag was previously unreachable (_resolve raised unconditionally) and scroll direction never left the tool boundary.
• Viewport — set_viewport, which makes the responsive checks in the attack catalog executable for the first time.
• Re-runnable suites — plans persist per repo with per-scenario history and run_plan, so "re-run my checkout suite after this change" is one call.
• Honest ledger — absence from the latest run is absent/not_run, not verified. Only an explicit sign-off is verified.

The bug worth reading the diff for

The new action types round-tripped into ReproStep but had no branch in replay_spec, so they counted as replayed without being performed. verify_fix could report a bug fixed while it was still there, and run_plan could report a scenario PASSED having done nothing. Fixed, with the invariant documented — plan.run_scenario and reverify.replay_spec share one executor, so any future action type must get a branch there.

Known follow-ups (not in this PR)

1. AnthropicDriver has no stop_reason handling. Opus 5 has thinking on by default and max_tokens caps thinking + response together; at 2048 a dense screen can truncate, degrade to "{}", and be parsed as a no-op step. A refusal does the same. The brain can go quietly blind mid-run.
2. get_adapter now raises when E2B_API_KEY is unset for sandboxed web/Electron — a deferred failure becomes a hard one at construction.
3. Driver default claude-sonnet-4-6 → claude-opus-5: ~1.7x per token, and a separate rate-limit bucket. Wants a CHANGELOG line.
4. Vendored axe-core has no NOTICE / third-party entry (MPL-2.0 header is intact in the file).
5. The E2B sandbox path still loads axe from the CDN — it reports axe_error now, but a CSP'd app there still gets no axe results.

🤖 Generated with Claude Code