feat: add calibration dashboard and history

[BLCCoreStudio]

Summary

Adds a localhost Calibration workspace over the existing multi-evaluator calibration core. It operates on actual saved evaluation artifacts plus workflow evaluator IDs, persists append-only calibration reports, and never rewrites original evaluations or workflow sidecars.

Dashboard

• dedicated /calibration workspace
• compatibility-grouped candidate selection
• two-or-more-evaluator calibration
• scalar, pairwise, and timestamped-audio agreement views
• evaluator aggregate bars
• criterion observation/agreement table
• audio pair agreement with F1, severity agreement, and temporal similarity
• configurable audio timestamp tolerance
• JSON export
• append-only calibration history with reopen/download support

Artifact and attribution boundaries

• candidates are read from workspace/evaluations/
• evaluator IDs are read from matching workflow sidecars
• a missing or corrupt workflow sidecar does not hide a valid evaluation; it is shown as unavailable for calibration because attribution cannot be established safely
• client-side compatibility grouping is UX only; the server and calibration core revalidate task/type/rubric/source compatibility
• reports are stored separately under workspace/calibrations/
• source evaluation artifacts remain byte-for-byte unchanged while generating a calibration report

Semantics

Calibration remains diagnostic. The dashboard does not:

• infer which evaluator is correct
• automatically pass/fail/rank evaluators
• define a universal agreement threshold
• adjudicate disagreements
• use an external LLM or telemetry
• upload evaluation or calibration data

Packaging and CI

• package version bumped to 0.6.0
• README and architecture updated to the real browser/dashboard surface
• dedicated Calibration Dashboard workflow checks dashboard backend tests, JavaScript syntax, and wheel-packaged HTML/CSS/JS assets
• full feature-branch CI is green on Python 3.11 / 3.12 / 3.13 with Ruff, strict mypy, 77 tests, 91.79% coverage, wheel/CLI smoke checks, live localhost HTTP/persistence checks, and desktop/mobile Chromium E2E regression coverage

Additional robustness fix

During edge-case testing, a corrupt workflow sidecar was found to hide an otherwise valid evaluation from calibration candidates. The backend now isolates those failures: the evaluation remains visible but is marked non-calibration-ready until evaluator attribution is valid.