fix(rca): normalize confidence score scale (no more "9000%")

[WZ]

Problem

The synthesis LLM is inconsistent about the confidenceScore scale — some completions emit a 0–1 fraction (0.9), others a 0–100 value (90 / 95.00). Stored unnormalized, every consumer that does × 100 rendered nonsense for the 0–100 case:

• investigation header: "9000%"
• report card: "(95.00)"
• the low-confidence styling (< 0.5) silently never fired for 0–100 scores

Found during browser-QA of the orchestrator; unrelated to that work, so it ships independently off main.

Fix

• New src/lib/confidence.ts — confidenceFraction (coerce either scale to a clamped 0–1) + confidencePercent (0–100 int). Shared web/server.
• Normalize at the source (synthesis step) so new reports persist a 0–1 fraction — also keeps the runner's low-confidence gate (which assumes 0–1) correct.
• Defensive at every display site so reports already persisted on the wrong scale still render right: RcaReport header + low-confidence banner/styling, InvestigationPane metadata, ChatPane card, markdown export, Slack notifier, runner event log.

Test

• New confidence.test.ts (both scales → same %, clamps pathological 9000, handles null/NaN).
• tsc clean, full suite 2409 passing, web build clean.