Connect OpenCode2 file inspections to local evidence specialist

[Noisemaker111]

Adds inspect_file to OpenCode2 Code Mode: the frontier supplies an exact known path and evidence intent, native read performs authorization and instruction discovery, and the unchanged local specialist selects verbatim source lines. The result includes original file line numbers, explicit page/failure fallback, and a raw page that can be reopened through native read. Python receives already-authorized text and never reads the target itself.

This is an opt-in file-page evidence integration. Routing the pilot's five command-planning operations remains separate. The plugin has been exercised in isolated launches of the configured host and has not been activated globally.

Validation: 26 Python tests, six Acorn parser tests, three adapter core checks, and diff checks pass. Real OpenCode2 beta-19398 with the configured gpt-5.6-sol/xhigh route and shell-specialist-pilot exercised discovery, Unicode/apostrophe paths, offset-correct evidence, saved raw reopening, native permission denial, page continuation, and context/selection failure fallback. Saved selected lines were reopened and compared with original fixtures.

Observed limitations are recorded in opencode/HOST_RESULTS.md: the model selected irrelevant INFO lines on long pages and failed both no-match trials; the frontier successfully completed the final answer using native fallback. One frontier run needed tool-name retries. Specialist call timings are separated from whole CLI sessions (61.4s and 129.9s for different multi-case workloads); no whole-workflow speedup or token/cost saving is claimed. Private logs and model artifacts are excluded.

Related: PR #3 appeared during verification and separately changes runtime capacity and empty-output handling. Its prediction API retains this adapter's call signature and defaults. This PR's nonempty no-match failures are a different observed case; no changes from #3 were incorporated.