fix(runtime): name the session model in the system prompt

[twtrubiks]

Summary

• The # Environment context section of the system prompt hardcodes Model family: Claude Opus 4.6 (FRONTIER_MODEL_NAME) for every Anthropic model. As a result claude-opus-5-5, claude-sonnet-4-6, and every other Claude model answer "I am Claude Opus 4.6" when asked which model they are, even though the request goes to the configured model.
• Fix: the Claude family label is now plain Claude, followed by a Model: <resolved model id> line. Non-Anthropic models keep the neutral an AI assistant label from Non-Anthropic providers inherit hardcoded Claude identity in system prompt #2822 and get their model ID line too.
• /model reused the system prompt built at startup, so after a switch the prompt would still name the previous model. set_model now rebuilds the system prompt for the new model.
• FRONTIER_MODEL_NAME is removed (it had no other users). load_system_prompt and load_system_prompt_with_context take the model ID as a new argument; all three callers (the CLI runtime, the system-prompt command, and sub-agents) already had it in scope.

 - Model family: Claude Opus 4.6     # before, for any Claude model

 - Model family: Claude              # after
 - Model: claude-opus-5-5

This is independent of #3355–#3358 and merges cleanly with each of them.

Anti-slop triage

• Classification: actionable-fix
• Evidence: repro below; new test renders_configured_model_instead_of_a_fixed_claude_version; updated renders_default_claude_model_family_identity, renders_generic_model_family_identity_without_claude_label, and subagent_system_prompt_uses_resolved_model_identity.
• Non-destructive review result: merge candidate

Reproduction

1. claw --model claude-opus-5-5 system-prompt | grep "Model family" prints Model family: Claude Opus 4.6.
2. claw --model claude-opus-5-5, ask "which model are you?" → the model says it is Claude Opus 4.6.

After the fix, system-prompt prints both lines above and the model identifies itself as claude-opus-5-5.

Verification

• scripts/fmt.sh --check
• cargo test -p runtime -p tools -p rusty-claude-cli: all pass except tools::tests::powershell_runs_via_stub_shell, which fails locally only because pwsh is not installed (same on main).
• cargo clippy -p runtime -p tools -p rusty-claude-cli --all-targets: no new warnings in the touched code.
• Manual, live Anthropic API: asked claude-opus-5-5 which model it is; it named claude-opus-5-5.
• Manual, /model: pointed ANTHROPIC_BASE_URL at a local server that records request bodies and returns a canned stream. In one REPL session I sent a prompt, ran /model claude-haiku-4-5, and sent another. The first request had model: claude-opus-5-5 and - Model: claude-opus-5-5; the second had model: claude-haiku-4-5 and - Model: claude-haiku-4-5. I checked the request bodies directly because asking the model to quote its own prompt was not reliable.
• Targeted tests/docs checks ran, or the gap is explicitly recorded.
• git diff --check passes.
• No live secrets, tokens, private logs, or unrelated generated churn are included.

Compatibility and known gaps

• runtime API: FRONTIER_MODEL_NAME is removed and load_system_prompt* gain a model parameter. All callers are inside the workspace and are updated here.
• Rebuilding the prompt on /model also re-reads git status and instruction files at that point, the same work done at startup. The prompt cache is invalidated by a model switch anyway.
• The # Runtime config section still embeds the raw settings JSON. If settings set "model" and --model overrides it, the model sees both values; it picks the active one from the Model: line. I left that section unchanged.

Resolution gate

• If this PR resolves an issue, the issue number and fix evidence are linked. (No existing issue; related to Non-Anthropic providers inherit hardcoded Claude identity in system prompt #2822, which made the label provider-aware but kept the hardcoded version.)
• If this PR should not merge, the rejection/defer rationale is evidence-backed and does not rely on vibes.
• I did not merge/close remote PRs or issues from an automation lane without owner approval.

[XinwangTao]

Fix looks correct: hardcoding Model family: Claude Opus 4.6 made every Claude model misidentify itself, and the resolved Model: <id> line plus the neutral label for non-Anthropic providers addresses that properly. Nice catch on /model reusing the startup-built prompt — rebuilding the system prompt on set_model is important so the identity updates after switching models. Test coverage for the three identity cases looks solid.

[XinwangTao]

这个修复很到位：直接在 system prompt 里写入实际解析的模型 ID，而不是硬编码固定的 Claude 版本，从根上解决了模型自报身份错误的问题。附带把 /model 切换后旧 prompt 残留的问题也一并修了，很关键。复现路径清晰、改动集中，测试也覆盖了默认/非 Anthropic 模型身份，支持合并。