Build the comprehensive V7 memory platform

[Da7-Tech]

Summary

• close the consolidated C-01 through C-34 defect register with regression coverage
• deliver roadmap A1 through A14: same-file MCP, transactional bulk ingest, persistent semantic sidecar, user/project tiers, deterministic journal merge, doctor/growth receipts, storage and privacy lifecycles, typed memory, policy hooks, deterministic single-file builds, governance files, and a public scoreboard
• preserve the stdlib-only offline default and the one-file distribution while moving development to ten deterministic source-domain fragments
• publish English/Arabic parity, truthful security boundaries, immutable benchmark manifests, generated claims, and eight raw evidence artifacts
• make Windows byte handling, exported paths, mutation staging, CRLF preservation, temporary cleanup, and journal digests portable across the full matrix

Verification

• 379 unit and acceptance tests pass with ResourceWarnings treated as errors
• deterministic builder, claims gate, workflow YAML parse, privacy scan, and diff checks pass
• all nine Ubuntu, macOS, and Windows cells pass on Python 3.9, 3.12, and 3.14
• retrieval: recall@1/5 = 1.00/1.00 at 100 and 1,000 nodes; multilingual 24/24; discrimination 12/12
• slot conflicts: 25 true positives and 25 true negatives, precision/recall 1.00
• fuzz: 300 graph plus 120 CLI cases, zero failures; 180-day soak passes
• five-year autonomy: archive rotation, bounded signals, one guard per agent file, green doctor, and exact growth accounting
• 10,000-fact transactional ingest uses one graph commit, one journal batch, and one signal batch, with a 109.2x conservative lower-bound speedup
• mutation analysis uses a green 376-test product baseline, explicitly excludes only the three self-referential evidence tests, and confirms every candidate kill in isolation: mind.py 43/120 killed with one parallel-noise reclassification; LongMemEval 35/120 killed; zero invalid, timeout, compile, or infrastructure outcomes
• LongMemEval 50-question subset: BM25 leads evidence metrics at 0.660/0.920; mind offline is 0.500/0.840; concept sidecar is 0.560/0.840. This subset does not measure graph, temporal, contradiction, privacy, merge, or autonomy capabilities and is not presented as an overall product ranking.

Compatibility And Risk

• development version is 7.0.0.dev0; the pinned stable 6.2.10 artifact and checksum remain unchanged
• old graph/journal formats remain readable; migrations, crash recovery, restore, and rollback paths are tested
• optional semantic modes are explicit process boundaries; offline behavior remains the default
• the change is intentionally a large preview; remote governance is enforced, all required checks pass, and open secret-scanning and dependency alerts are empty