docs(changelog): document parse/validate retry fix and tooling changes

The earlier Unreleased entry only covered the Supervisor evaluation
loop feature. This adds the companion Fixed and Changed sections for
the same PR:

- Fixed: spurious first-call INVALID_JSON_RESPONSE failures via the
  new parseAgentJson extractor and runAgentWithValidation retry
  wrapper.
- Changed: all 4 invokers now route through runAgentWithValidation;
  FileStore and SQLiteStore accept a storage path via constructor
  (default unchanged) to unblock integration tests.
- Added: note about the first test suite and CI wiring.

Author: iamvirul

CHANGELOG.md

@@ -12,6 +12,17 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - `COUNCIL_EVAL_RETRIES` env var — clamped to `[0, 5]`. Set to `0` to restore pre-0.5 advisory-only behaviour. Non-integer values fall back to the default.
 - Retry count recorded in `session.metrics.eval_retries`, visible via `get_council_state` and in the result summary footer.
 - `supervisor_feedback` field added to `AgentInvokeOptions`; Executor and Aide prompts render a clearly-delimited `--- SUPERVISOR FEEDBACK ---` block when set.
+- First test suite for the repo — 160 tests across 15 files covering config, schemas, stores, agent invokers, the feedback formatter, the eval loop, orchestrator routing, and integration tests for FileStore and SQLiteStore against real temp dirs.
+- CI now runs `npm test` on every push (new `Test` job in `ci.yml`) and as part of PR validation (`pr-check.yml`).
+
+### Fixed
+- **Spurious first-call `INVALID_JSON_RESPONSE` failures** — the claude CLI occasionally emits JSON wrapped in prose ("Here's the output: {...} Let me know!") or produces a single flaky sample (missing field, wrong enum) on the first try. A strict `JSON.parse` + schema check rejected both, surfacing as `INVALID_JSON_RESPONSE` before the Supervisor evaluation loop ever ran. Fixed with two new layers:
+  - `parseAgentJson` extracts JSON via fence-match → balanced-brace extraction (string-aware, so braces inside JSON string values are not mis-counted) → raw trimmed, returning the first candidate that parses.
+  - `runAgentWithValidation` wraps `runAgent` + parse + Zod validation with a one-shot retry on parse/validate failure. Transient flakiness recovers silently; CLI transport errors still propagate unchanged.
+
+### Changed
+- All 4 agent invokers (Chancellor, Executor, Aide, Supervisor) now route through `runAgentWithValidation` — removes ~80 lines of duplicated fence-match + parse + validate boilerplate. External behaviour is unchanged on success; failures are rarer and still produce the same `CouncilError` code (`INVALID_JSON_RESPONSE` for Chancellor/Executor/Aide, `SUPERVISOR_ERROR` for Supervisor) on final give-up.
+- `FileStore` and `SQLiteStore` accept a storage path via constructor argument. Default is still `~/.council/sessions` and `~/.council/council.db` respectively; the parameter exists so integration tests can use temp dirs instead of polluting the user's home directory.
 
 ## [0.4.0] - 2026-04-20