Track A NHI re-run + in-campaign fix batch.

Focused re-run of the NHI test playbook Track A against the post-PR-#820 binary, plus complete remediation of every defect surfaced during the re-run — all landed in v0.7.0, no deferral. Per the prime directive testing-loop addendum.

SHIP · 85 PASS / 0 FAIL across 12 phases

In-campaign fixes

10 GitHub issues closed with retest + re-check evidence.

#826memory_update MCP coverage (via #859) #829verbose token budget 15570 → 9507 #830TTL "extend" docs drift #831memory_promote step-skip docs drift #837NHI-P2 forget scratch-DB discipline #859MCP tools/list optional schema restored #860get_links surfaces temporal+attest #861archive_list preserves metadata + tags JSON #862tool count 70+1 disambiguation #865playbook boot-directive worktree drift

Files in this campaign

Read in this order.

README.mdCampaign index, hard rules, namespace convention, provenance→ track-a-nhi-results.mdTrack A raw results, all 12 phases + verdict (engineering)→

Three audiences

Same campaign, three abstraction levels.

Every campaign produces three audience-facing pages plus the engineering writeup. Per operator directive 2026-05-18 pm.

For non-technical readersWhat ai-memory does, what we tested, what the verdict means — plain English, zero jargon→ For executives / PMs / decision-makersVerdict, risk profile, cost, vector-DB comparison, v0.7.0 vs v0.8.0 scope, honest disclosure→ For SME engineers + architectsReproducibility, methodology, token budget, schema ladder, coverage, refactor LOC, security/code verdicts, six-level provenance maturity→

Discipline

Testing-loop addendum compliance.

No deferral. Every issue surfaced during the re-run was fixed in v0.7.0.
Banned framings refused. "non-blocking", "P2/P3 follow-up", "surface-level", "vN+1 polish" disallowed.
Audit trail mandatory. Every GH issue body links to ai-memory evidence; ai-memory evidence links to GH issue id; commit messages reference both.
Retest discipline. Each fix verified via the same scenario that surfaced it (retest) PLUS a fresh probe from a different angle (re-check). No close-as-fixed without both green.
Recompile + batch retest. The campaign rebuilt the release binary once after all fixes landed, then ran the batch retest sweep against the new binary via CLI + raw MCP probes.

Provenance

Reproducibility contract.

Campaign date: 2026-05-18 (morning re-run + afternoon fix batch)
Operator: justin@alpha-one.mobi
Authoring agent: Claude (Opus 4.7 1M context)
Binary at retest: git SHA e99fb0e on branch local/install-815-816
DB: operator's live MCP DB, schema v43
Models: nomic-ai/nomic-embed-text-v1.5, cross-encoder/ms-marco-MiniLM-L-6-v2, gemma4:e4b. Tier: autonomous.
Prime directive: memory f1dca8fa-6c33-4139-b0b5-389cca45b921 (testing-loop addendum)

Other campaigns

10 GitHub issues closed with retest + re-check evidence.

Read in this order.

Same campaign, three abstraction levels.

Testing-loop addendum compliance.

Reproducibility contract.

Related runs.