feat(opencode): inject x-opencode-session header for prompt caching

[pspranger-throw]

Send the x-opencode-session header on every API request to OpenCode Go/Zen providers so the proxy can cache the prompt prefix across requests in the same conversation.

The header is injected per-request in _build_api_kwargs() across all three API modes:

• chat_completions (opencode-go default, opencode-zen default)
• anthropic_messages (opencode-go minimax-*, opencode-zen claude-*)
• codex_responses (opencode-zen gpt-*)

Includes 9 tests covering all three paths, negative cases (no header leakage to other providers), and state mutation (model switch, session rotation after compression).

What does this PR do?

Adds a header to all opencode zen/go api calls to ensure that opencode can track sessions for context caching, otherwise each call pays for the full uncached context!

Solution orients itself after the existing x-grok-conv-id implementation.

Related Issue

Related to #5252. This is a minimal fix for OpenCode Go and Zen while the broader routing table architecture from #5252 is developed.

This PR takes a minimal, targeted approach: direct conditionals in
_build_api_kwargs() for the OpenCode provider only. When the
data-driven routing table from #5252 lands, the OpenCode entries
can be migrated into it.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

**`run_agent.py` -- `_build_api_kwargs()` (3 injection points)**

- **`chat_completions` path** (~line 7168): Inject `x-opencode-session` header into `api_kwargs["extra_headers"]` when `"opencode.ai"` is detected in `self._base_url_lower` and `self.session_id` is set.

- **`anthropic_messages` path** (~line 6860): Capture the return value of `build_anthropic_kwargs()` into a local `kwargs` variable (previously returned directly), then inject the header before returning. This route is used by `opencode-go` with `minimax-*` models and `opencode-zen` with `claude-*` models.

- **`codex_responses` path** (~line 6957): Inject the header alongside the existing `x-grok-conv-id` injection for xAI. This route is used by `opencode-zen` with `gpt-*` models.

The header is injected per-request, reading `self.session_id` and `self._base_url_lower` fresh on every API call. This ensures it automatically follows runtime state changes from model switches (`/model`) and context compression (which rotates `session_id`).

Defensive attribute access: Uses getattr(self, "_base_url_lower", "") rather than
self._base_url_lower because some code paths (notably tests using object.__new__()
to bypass __init__) reach _build_api_kwargs without the setter having run. Returns
"" which safely skips the opencode check — correct behavior for non-opencode agents.

**`tests/run_agent/test_run_agent.py` -- `TestOpencodeSessionHeader` (9 tests)**

- 3 positive tests: header present with correct value for `chat_completions`, `anthropic_messages`, and `codex_responses` API modes
- 3 negative tests: no header when session_id is None, no header for non-OpenCode providers (OpenRouter, Anthropic, OpenAI), no header leakage across API modes
- 2 state mutation tests: header follows `base_url` changes (model switch scenario) and `session_id` changes (context compression scenario)
- 1 edge case: header uses child session_id after compression, not stale parent

How to Test

1. Used my opencode-go subscription, verified that sessions now cache properly on my dashboard (actual entries in session columns and pricing staying low with growing context)
2. Went through a compression and verified new session ID after compression as base for new caching
3. Cannot test zen provider though -> only have opencode go

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate (-> see feat(routing): provider-aware session affinity hints for prompt caching #5252 )
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass ( there are pre-existing unrelated failures!)
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: cachyOS

Documentation & Housekeeping

• N/A - I've updated relevant documentation (README, docs/, docstrings) — or N/A
• N/A - I've updated cli-config.yaml.example if I added/changed config keys — or N/A
• N/A - I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — or N/A
• N/A - I've considered cross-platform impact (Windows, macOS) per the compatibility guide — or N/A
• N/A - I've updated tool descriptions/schemas if I changed tool behavior — or N/A

Screenshots / Logs

[pspranger-throw]

@kovyrin @jbarket @chalkers @leon

Requesting review please or adding of other reviewers please, since I cannot add reviewers myself from outside fork.