fix(config): keep startup LLM fallback in-memory only (#3229)

[ilblackdragon]

Closes #3229.

Summary

• A transient hydration failure on startup (DB read race, secrets decryption hiccup) used to overwrite the user's persisted llm_backend with nearai and delete selected_model from the DB, turning a one-off fallback into a permanent reversion of the configured provider. After the fix, the fallback stays in-memory: the runtime still demotes to NearAI to keep the instance usable (Issue: Incomplete Provider Configuration Causes Instance Startup Failure (Non-TEE) #2514 crash-loop prevention is preserved), but the user's DB row is left untouched, so the next restart with credentials available picks up their original config.
• Removed the post-fallback DB write block in Config::resolve_llm_with_secrets_inner, plus the now-unused fallback_fired / normalize_backend helpers and their five unit tests (those existed solely to gate the deleted write). Net -88 lines on the bug fix.
• Updated the stale comment in LlmConfig::resolve_nearai_fallback that referenced the removed DB sync.

Scope expansion: workflow brittleness fix

While debugging this PR's CI, GitHub failed to publish refs/pull/3324/head (confirmed by git ls-remote against origin — only pull/3324/merge exists; comparable open PRs have both refs). The regression-test-check.yml workflow was hard-failing on git fetch origin pull/N/head:pr-head before any of its actual checks could run. This was a pre-existing brittleness — the workflow assumes refs/pull/N/head is always published, which GitHub does not guarantee.

Bundled fix: derive the head SHA from github.event.pull_request.head.sha (always present in the event payload) instead of fetching refs/pull/N/head. The merge commit actions/checkout already produces with fetch-depth: 0 has the head SHA as its second parent, so it's locally reachable. Future PRs hit by the same GitHub ref-publishing bug will not need this manual unsticking.

The empty ci: re-trigger workflows commit in the history is residue from earlier debugging — it pushed a refreshed SHA to nudge GitHub into publishing the missing ref (didn't work, hence the workflow change).

Test plan

• cargo fmt
• cargo clippy --all --benches --tests --examples --all-features — zero warnings
• cargo test --lib — 5614 passed
• New regression test config::tests::startup_fallback_must_not_overwrite_persisted_user_backend — seeds llm_backend = "openrouter" + selected_model = "openai/gpt-4o-mini", calls the startup path with no secrets store (forces the unusable-config fallback), asserts the runtime falls back to NearAI and the DB row is unchanged. Would have failed on main (the row got rewritten to nearai and selected_model deleted).
• scripts/pre-commit-safety.sh passes
• Workflow change verified in CI on this PR (the very thing it's trying to fix)

🤖 Generated with Claude Code

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[Copilot]

Pull request overview

Prevents the startup-time “fallback to NearAI” safety mechanism from permanently overwriting a user’s persisted LLM provider/model in the DB when secrets/config hydration is temporarily unavailable, while keeping the crash-loop prevention behavior (#2514) intact.

Changes:

• Removed the post-fallback DB synchronization in Config::resolve_llm_with_secrets_inner so the fallback remains in-memory only (#3229).
• Updated NearAI fallback commentary in LlmConfig::resolve_nearai_fallback to reflect the new persistence behavior.
• Adjusted the regression-test enforcement workflow to use the PR head SHA directly instead of fetching refs/pull/N/head.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
src/config/mod.rs	Stops persisting fallback backend/model into DB; adds regression test ensuring DB row is preserved on startup fallback.
src/config/llm.rs	Updates fallback comments to reflect in-memory-only behavior for selected_model.
.github/workflows/regression-test-check.yml	Makes the workflow more robust by comparing against the PR head SHA from the event payload.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.