# M01–M10 implementation evidence

2026-10-13. This is engineering evidence, not production sign-off.

## Completed locally

- M01: Acquired official project HTML/PDF originals and area references; created 6 draft documents with 20 facts, including 12 FAQ, in English/Chinese/Bahasa Malaysia. SHA-256 manifest retained. Business review and usage permission remain pending mentor/Wai Kit.
- M02: Prepared 20 retrieval cases, 30 end-to-end candidates and all 12 original QA questions, plus additive schema 1.0 contract. The report's missing exact Chinese input remains missing.
- M03/M04: Site migrations, offline queue import, explicitly configured embedding adapter and Qdrant client, immutable versions, ready release activation/rollback, withdrawal and applicability/entity filtering. Default flag is off. No real provider key was used.
- M05: Explicit budget/quotation/area distinction, region/project aliases, guide camera focus, revision-locked brief updates, idempotent turns, cancellation and stale-revision behavior. Ambiguous price ranges and foreign currencies are not silently converted.
- M06/M07: Admin-gated knowledge planner tool, snapshot-independent facts, mixed result composition and citation namespaces, reviewed translated knowledge claims, deterministic numerical display templates, unknown citations/placeholder clarifications rejected.
- M08: Document source drawer, saved version/excerpt/date display, knowledge-only map/selection preservation, existing chart reuse and rate-limit waiting guidance.
- M09/M10: Targeted regression, client/SSR builds, runbook and fixed pilot/feedback/recovery steps prepared.

## Verification

Implementation worktree: PHP 8.4.21; isolated local MySQL database `petav3_map_rag_20261013_testing`, including catalogue test aliases. No production or business database was migrated. Laravel's test guard checked the loopback hosts and `_testing` database names before boot.

- PHP PropertyLab unit suite + Advisor feature suite: **70 tests, 589 assertions passed** on the implementation baseline. After integration into the earlier local dev-chen, its applicable suite passed **66 tests, 453 assertions**; the count differs because unrelated newer tab/Explorer tests were not imported.
- JavaScript PropertyLab composables/components/utils: **75 tests across 15 files passed**.
- Vite client build and SSR build passed.
- PHP lint and `git diff --check` passed; JSON material/eval/contract files parsed successfully.
- Fixed positive cases checked with isolated, explicitly labelled test approvals and HTTP/model fakes. This verifies content wiring, filters, persistence and source binding; **it is not a measurement of real model selection or embedding recall**.
- Existing PHPUnit XML uses a deprecated schema; this is a runner warning, not a failing test. Existing build warnings include large chunks and unresolved runtime image URLs. No unrelated schema/asset/lockfile cleanup was included.

Manual review covered the default-off/admin boundary, SQL connection separation, index authority recheck, immutable citations, no raw private prompt logging, explicit service configuration, partial-vs-unavailable behavior, and fail-closed project mapping. Map/data bridging additionally requires a separate coordinate/mapping review in the published knowledge version and active LINK/confidence/decision metadata in the catalogue.

## Not yet accepted

1. Mentor/Wai Kit review of facts, translations and use scope, and the three project mapping/coordinate sign-offs. Draft packages remain rejected by the importer.
2. G01/G02 production release, actual Research model, master/COPY identity, complete snapshot/coverage and original-file evidence.
3. Approved embedding and Qdrant service endpoints/credentials, real indexing/rebuild and multilingual recall evaluation. No infrastructure was deployed or purchased.
4. Real 30-case model evaluation, full 12-question QA replay with missing inputs resolved, browser journey on verified map data, five concurrent sessions, P95 latency and actual costs.
5. Approved staging installation and 3–5 human pilot reviewers. No production deployment, invitation, permission expansion or Git push occurred.

Therefore M01 business approval and M09/M10 live/pilot gates are **pending**, even though the local implementation and delivery preparation are present.

## Local integration

Only the three task commits were cherry-picked into local dev-chen: 2871b3a66 (materials/contract), d3ccd08b2 (knowledge backend), 00d852eb4 (Advisor integration). A recoverable backup `codex/backup-dev-chen-rag-20261013` retains the previous 5bba44e0b head. One test-file insertion conflict was resolved by retaining the target's existing helpers and inserting only the two new task regression methods; no unrelated upstream helper/tests were included.

The integrated source passed its 66 PHP tests and 75 JavaScript tests. Client and SSR builds were also repeated on the exact integrated source in the isolated worktree with freshly installed locked dependencies. No push or deployment was performed.

## One-command deployment preparation

2026-10-13: added a reviewed bootstrap for the existing master deployment, plus `propertylab:advisor-prepare` read-only preflight, synchronous import/index smoke checks, activation and guarded recovery. Research/Knowledge remain default-off; the command never edits server flags or approves documents. The newer Advisor audience lock must retain the admin preview.

- Earlier local dev-chen baseline: **72 PHP tests / 503 assertions passed** in the same isolated loopback test database, with fake embedding/index/model responses. Six new command tests cover draft/hash/site/DB/model/mapping/chart/audience/report refusal, no HTTP during preflight, index failure retaining activation, maintenance, repeat install, recovery and refusal to overwrite a later activation.
- Bootstrap shell checks passed for fetching the merged script on first install, expected SHA/flags, wrong origin, dirty checkout, privilege/lock checks and failure exit propagation. These are fake CLI checks, not a Linux service deployment.
- PHP syntax, Bash syntax and `git diff --check` passed. The temporary runner needs a 512 MB PHP limit; the repository's 128 MB default was insufficient during route bootstrap. The deprecated PHPUnit XML warning remains.
- Manual review covered configuration-only model metadata, exact approved package bytes, sanitized errors, database identity/read connection checks, synchronous activation after semantic probes, and failure recovery. Real approval, recall/model QA, user credits, map journey, costs and server recovery remain pending in DEPLOYMENT.md.

Remote dev-chen was observed at d788451728c7536c542eafb3d825b29e8df18a95, 405 commits beyond the earlier common baseline; remote master was observed at f3f2d4250243cace34e4e2ada1fe5ff5011dee9f. The clean local dev-chen was backed up as `codex/backup-dev-chen-before-advisor-deploy-20261013` (b9fb57e84), then synchronized with origin/dev-chen by a clean merge preserving all four existing task commits. Only the new task commit was cherry-picked from the deployment branch: f80689a8a. Its source tree exactly matched the tested deployment checkout.

- Synchronized dev-chen source: **76 PHP tests / 639 assertions**, **75 JavaScript tests / 15 files**, client and SSR builds passed after fresh locked npm installation. Bash/CLI bootstrap checks also passed.
- Master merge preview: `codex/map-advisor-master-preview`, 3b2fdec51, merged the observed master into the deployment checkout without conflicts. **77 PHP tests / 712 assertions**, **75 JavaScript tests / 15 files**, client and SSR builds passed. The extra PHP case comes from master. Composer/npm lockfiles were unchanged between synchronized dev-chen and this preview.
- Master had 84 commits absent from remote dev-chen. They remain on master; their full release/migration scope must be reviewed before the real deployment. The preview checks the Advisor area and builds, not every business feature.

These passes use test approvals and fake provider/index responses, not real deployment approvals or retrieval/model quality measurements. Production SHA/model/database identities, approved materials/services and staging acceptance remain unconfirmed. No push, PR creation or production execution occurred.

## Herd local verification

2026-10-13: the existing `petav3.test` site was repaired and exercised in a browser. The actual Herd checkout passed **76 PHP tests / 639 assertions** on isolated testing connections with Herd PHP 8.4.21. Matching-source 75 frontend tests, client/SSR builds and bootstrap checks passed. Admin preview, Mapbox controls, independent brief inputs and save/reload persistence worked. Real snapshot/chart data, AI/RAG services and authoritative knowledge were unavailable; no real generation request was sent. See [HERD_TEST.md](HERD_TEST.md) for precise evidence, local changes and remaining gates.
