fix(claude): preserve model selector provenance

[acedatacloud-dev]

Problem

When a Claude Code transcript is resumed through Coding Bridge, the transcript's resolved model ID (for example claude-opus-5) was reused as the launch selector. Selector-only options such as opus[1m] are not present in the transcript, so cross-device resume silently changed the context profile and triggered compaction near the ordinary context threshold.

Root fix

• Version the sidecar model contract and persist exact model_selector provenance separately.
• Expose model_selector, resolved_model, and model_contract=2 as distinct protocol fields.
• Treat legacy model sidecars as ambiguous instead of guessing a model family or fabricating [1m].
• Omit an explicit model override when selector provenance is unknown, allowing Claude Code settings to remain authoritative.
• Log requested selector and provider-resolved model separately at launch/init.

No token thresholds, title/session special cases, or forced 1M selector are used.

Verification

• python3 -m pytest tests/test_session_meta.py tests/test_session.py tests/test_connection.py tests/test_history.py tests/test_claude_entrypoint.py tests/test_resume_guard.py — 86 passed
• ruff check ... on all touched Python/test files — passed
• Full local suite: all affected tests pass; unrelated pre-existing environment failures remain in channel log-capture/portal tests and one live Codex capability assertion.

🤖 对抗评审 (reviewer: codex, 1 round)

• Reviewer launch was attempted read-only, but the local Codex client hung after its model-catalog endpoint returned an incompatible schema. No findings were produced; focused contract review and cross-repo tests were completed manually.

🤖 Generated with Claude Code