fix(analyst): keep Python and TypeScript finding validation identical

[drewstone]

Defect

A paid DSPy RLM run can return findings that the Python bridge accepts and the TypeScript boundary silently drops.
The Python success prevents the bounded format-repair turn from running.
A consumer with requireStructuredFindings: true then fails after useful work completed.

Observed with @tangle-network/agent-eval@0.145.11 through Braid 0.1.2 and a retained Claude Code CLI Bridge run.
The model submitted four structured rows.
All four used free-form subjects such as What changed.
Python reported them as findings, while RawAnalystFindingSchema rejected every row.

Minimal reproduction

Python accepts any bounded non-empty subject in clients/python/src/agent_eval_rpc/dspy_rlm_bridge.py::_validate_finding:

row = {
    "severity": "info",
    "claim": "x",
    "subject": "What changed",
    "confidence": 0.9,
    "evidence": [{"uri": "trace://run/span/span", "excerpt": "exact quote"}],
}
assert dspy_rlm_bridge._parse_findings_json(json.dumps([row])) == [row]

The published TypeScript parser rejects the same row:

import { parseRawFinding } from '@tangle-network/agent-eval/analyst'

parseRawFinding({
  severity: 'info',
  claim: 'x',
  subject: 'What changed',
  confidence: 0.9,
  evidence: [{ uri: 'trace://run/span/span', excerpt: 'exact quote' }],
}) // null

parseFindingSubject('What changed') also returns null.

Production evidence

• Source trace: run-721d5d04-e603-4a3d-b05b-2ab66e7e2682
• Analyst model calls: 3
• Model output: four structured findings plus a prose answer
• Consumer result: zero findings and trace analyst 'question' returned no valid structured findings
• Provider usage: 6 input tokens, 2,921 output tokens, $0.6197 reported cost, 60,163 ms model latency

Required behavior

• Define one canonical raw-finding acceptance contract across the Python and TypeScript boundaries.
• Reject or repair every Python row that TypeScript would reject before Python reports success.
• Preserve per-row rejection reasons in the runtime result.
• Add a parity test with valid subjects, invalid subjects, unknown fields, bad evidence shapes, and non-finite confidence.
• Prove one malformed sibling does not erase valid siblings.
• Prove an all-invalid final submission triggers the one declared repair turn or fails with the exact row defects.

Store-dependent evidence resolution can remain in TypeScript.
The cross-language shape contract must not disagree before that stage.