[dcaldr]

Summary

Closes #1792

Why this change is needed

In CodeBurn, userMessage is load-bearing beyond display:

• Activity Classification: classifyConversation and refineByKeywords inspect turn.userMessage to categorize turns into Coding, Planning, and Exploration. Without prompt text, Antigravity turns without direct file edits collapsed into generic Conversation, misattributing ~60% of coding work.
• Pull Request Links: extractPrUrlsFromText(call.userMessage) scans user prompts for PR URLs. Antigravity previously attributed 0 PR links from prompt text.
• Session Identification: codeburn sessions and dashboards showed opaque fallback cascade IDs instead of task descriptions.

What this PR does

• Extracts the initial human prompt from brain/<cascadeId>/.system_generated/logs/transcript.jsonl (unwrapping <USER_REQUEST> and bounding to 500 characters, matching CodeBurn's cross-provider convention).
• Adds a defensive fallback to decode prompt text from steps (step_payload field 19.2) in SQLite when transcript logs are unavailable.
• Bumps ANTIGRAVITY_CACHE_VERSION = 10 and PROVIDER_PARSE_VERSIONS.antigravity (-user-msg-v1) so cached sessions gain prompt attribution with byte-for-byte cost and token parity ($0.00 drift).

Follow-up to precursor Antigravity PRs #1456, #1620, and #1712, following the precedent established in #1581 (ZCode prompt attribution).

Testing / Verification

• I have tested this locally against real data (not just unit tests)
• npm test passes
• npm run build succeeds
• Not a UI change

Verification Proof

Unit test suite (npx vitest run tests/providers/antigravity.test.ts):

 ✓ tests/providers/antigravity.test.ts (67 tests) 1043ms
 Test Files  1 passed (1)
      Tests  67 passed (67)

Real data verification (codeburn today --provider antigravity):

Before (Baseline - no prompt text, Coding collapsed into Conversation):

╭──────────────────────────────────────────────────────────────────────────────╮
│ By Activity                                                                  │
│                                                            cost turns 1-shot │
│ ██████████ Conversation                                   $7.49   629      - │
│ █░░░░░░░░░ Planning                                      $0.379    34      - │
│ ░░░░░░░░░░ Exploration                                  $0.0079     1      - │
╰──────────────────────────────────────────────────────────────────────────────╯

After (With Patch - 209 coding turns and 15 exploration turns properly categorized):

╭──────────────────────────────────────────────────────────────────────────────╮
│ By Activity                                                                  │
│                                                            cost turns 1-shot │
│ ██████████ Conversation                                   $4.88   402      - │
│ █████░░░░░ Coding                                         $2.22   209      - │
│ █░░░░░░░░░ Planning                                      $0.379    34      - │
│ █░░░░░░░░░ Exploration                                   $0.351    16      - │
╰──────────────────────────────────────────────────────────────────────────────╯

Zero Drift Accounting:

• Total spend: $7.88 (Before) vs. $7.88 (After) -> $0.00 drift.
• Total calls: 664 (Before) vs. 664 (After) -> 0 call drift.
• 227 turns shifted cleanly from generic conversation to verified coding and exploration.