fix(cli): preserve HTTP 400 details

[koriyoshi2041]

Summary

• classify HTTP 400 responses before timeout keywords in the server body
• preserve the actionable server response for rejected requests
• add a regression test for the batch-mode message from memory retain prints "Request timed out" (0s) for an HTTP 400 batch-mode rejection — swallows the self-explanatory error body #2912

Tests

• cargo test --manifest-path hindsight-cli/Cargo.toml --bin hindsight (70 passed)
• git diff --check

The repository-wide formatting check currently reports pre-existing formatting drift outside this change.

Closes #2912

[koriyoshi2041]

The generated-files failure appears transient or environment-specific. On the exact PR head 3fc58621a, I ran:

./scripts/generate-openapi.sh
./scripts/generate-docs-skill.sh
git status --short

The working tree stayed clean. CI reported only a 2-line change in skills/hindsight-docs/references/openapi.json, while the source and generated copy are byte-identical locally. I don't have permission to rerun the repository workflow; a rerun should distinguish nondeterministic generation from a persistent Linux-only difference.

[koriyoshi2041]

Correction after reproducing from a fresh detached worktree at exact head 3fc58621a:

The failure is deterministic and comes from the current base, not this CLI patch. The two committed OpenAPI copies already differ before generation:

-    "version": "0.8.5"
+    "version": "0.8.4"

hindsight-docs/static/openapi.json is 0.8.5, while skills/hindsight-docs/references/openapi.json is 0.8.4. Current main at 4dc834834 has the same mismatch. generate-docs-skill.sh copies the 0.8.5 source into the skill, producing the CI job's exact one-line diff.

So a rerun alone will remain red until the generated skill copy is refreshed on main or this branch is updated after that base fix. I am not adding the unrelated generated-file change to this focused CLI PR unless maintainers prefer that.