chore: post-PR-79 refinements for Vertex AI and curation parser

[yvonnedevlinrh]

Context

PR #79 fixed three production bugs (#76, #77, #78) in Vertex AI curation reliability. During review, several non-blocking quality refinements were identified that should be addressed in a follow-up.

Tasks

1. Extract Vertex AI magic numbers to named constants

File: llm/vertex.go

Timeout: 300 * time.Second (line 85) and MaxTokens: 16000 (line 119) are inline literals. The existing codebase defines retry-related values as named constants (vertexSynthMaxRetries, vertexSynthBaseDelay, vertexSynthMaxDelay). These two should follow the same pattern:

const (
    // vertexSynthTimeout is the HTTP client timeout for Vertex AI requests.
    // Large curation prompts (36K+ tokens) need ~120-180s for response generation.
    vertexSynthTimeout   = 300 * time.Second
    // vertexSynthMaxTokens is the maximum output tokens for Vertex AI responses.
    // Curation extractions routinely exceed 4K tokens; 16K provides headroom.
    vertexSynthMaxTokens = 16000
)

Convention: CS-007 (magic numbers should be named constants)

2. Update ParseExtractionResponse doc comment

File: curate/curate.go, lines 319-321
The doc comment doesn't mention that individual unparseable items are silently skipped. A caller reading only the doc would expect any parse failure to return an error. Add a line:
// Silently skips individual items that cannot be parsed after fallback attempts.
Convention: DR-001 (GoDoc accuracy on exported functions)

3. Strengthen regression test assertions

File: curate/curate_test.go, TestParseExtractionResponse_QualityFlagsString (line 648)

The test verifies item count and QualityFlags length but doesn't assert that other fields survived the delete + json.Marshal + json.Unmarshal round-trip. Add:

if files[0].Content != "Use OAuth2." {
    t.Errorf("files[0].Content = %q, want %q", files[0].Content, "Use OAuth2.")
}
if files[0].Confidence != "high" {
    t.Errorf("files[0].Confidence = %q, want %q", files[0].Confidence, "high")
}

Convention: TC-006 (regression test completeness)

Out of scope (informational, no action needed)

• Fallback scope: The fallback path deletes quality_flags on any unmarshal failure, not just quality_flags-specific failures. This is intentionally lenient for LLM output parsing — items that can't be fixed by removing quality_flags are correctly skipped via continue. No change needed.
• Unreachable error check: json.Marshal(fields) after a successful json.Unmarshal into the same type cannot fail. The defensive guard is harmless. No change needed.
• Global MaxTokens: MaxTokens: 16000 applies to both curation and compile callers. Compile outputs are typically smaller but the higher limit has no functional impact. No change needed unless per-caller configuration is added later.

####Scope

• Files: llm/vertex.go (2 lines), curate/curate.go (1 line), curate/curate_test.go (4 lines)
• Risk: None — pure refactor + doc + test strengthening, no behavioral change
• Labels: chore, good-first-issue
• Origin: PR fix(curate): improve Vertex AI reliability for large curation extractions #79 review findings