release: bump version to 0.6.1 (#117)

Changelog for 0.6.1 already landed in #115. This bumps the package
metadata so installs from PyPI (and `importlib.metadata.version`)
resolve to the correct release tag.

Co-authored-by: Gradata <[email protected]>

test(#109): update test_4_failure_detectors_exist to canonical module

After 5b47d92 stripped detect_* functions from quality_monitoring.py,
the SPEC compliance test asserted their presence on the wrong module.
Pointing it at scoring/failure_detectors where they live canonically.

Verified: test passes. Other failures in the full suite
(test_agent_precontext_*, test_cli_export_*, test_runner_invokes_*)
are pre-existing environmental flakes (live brain DB drift and Windows
cp1252 encoding) unrelated to this branch.

Co-Authored-By: Gradata <[email protected]>

chore: bump version to 0.6.0 for PyPI release

Co-Authored-By: Gradata <[email protected]>

v0.5.0: self-healing, scoped brains, adapters, distribution, correcti…

…on hardening

Cumulative release since v0.4.0 (2026-04-07):

- #66 platform_source auto-detect on correction events
- #67 fingerprint-based observation dedup
- #68 per-call rule injection hook with draft-relevance ranking
- #69 AST-aware severity classifier with edit-distance fallback
- #70 background worker queue with job dedup + graceful shutdown
- #73 differential-privacy scaffold (off-by-default)
- #77 close self-healing loop, wire into brain.correct()
- #78 Brain.scope(domain) + sub-agent inheritance
- #79 Mem0 memory adapter with push/pull/reconcile protocol
- #80 @gradata/cli npm + Docker daemon image + tag-triggered publish
- #81 close rule-to-hook auto-promotion in injection paths
- #85 correction hardening: provenance hash, adversarial blocklist, semantic diff
- #88 wire dormant rules.injected emit

Unblocks gradata-cloud so it can pin to PyPI instead of a SHA tarball.

fix: sdk-release workflow — remove stale sdk/ working-directory

The project root IS the SDK; no sdk/ subdirectory exists.

Co-Authored-By: Gradata <[email protected]>

S79: v1-slim — cut 19.6K lines, add brain.prove() + export_rules() + …

…10K stress tests

Core cut: removed 64 commodity files (patterns, integrations, cloud, sidecar,
benchmarks, advanced enhancements). Kept 55 files / 15.3K lines of core IP:
graduation pipeline, auto-correct hook, meta-rules, FSRS confidence, brain engine.

New features:
- brain.prove(): Wilcoxon signed-rank proof that corrections decrease post-graduation
- brain.export_rules(): 5 formats (skill/claude/cursor/system/json)
- 10K brain-session stress test (20 job-domain personas x 500 sessions)
- LLM pipeline stress test (55 text pairs, 11 domains, real diff_engine)
- Live brain stress test (360 real brain.correct() calls)

508 tests pass, 0 fail. 34.9K lines -> 15.3K lines (56% reduction).

Co-Authored-By: Gradata <[email protected]>

Fix loop-state.md path in session-start to use absolute brain vault path

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>