# Catalogue recommendations and multi-turn continuity acceptance

Date: 2026-10-16. Target: authenticated local Herd `petav3.test`, existing cloud AI access. Delivery: PR #210, `dev-chen → master`. No merge or production deployment.

## Data and implementation

The earlier comparison returned synthetic Research schemes and forgot a workplace correction. New purchase-project tools reuse the company's existing `catalogue` connection and published, site-visible Malaysian new-project records. They do not copy or scrape WK's coach implementation. WK displays PETA/PropertySifu project/layout snapshots; the local company catalogue already holds those real projects and source identities, so this change reuses that existing read path.

The local COPY's accepted PropertySifu records were collected **2026-07-10**, earlier than WK's October snapshot. This acceptance does not claim to have synchronized that newer snapshot. Production uses its existing master/COPY configuration. No new database, migration, environment variable, dependency or catalogue write is required by this follow-up.

Price, bedrooms, size, source/date and layout identity come from the same active source contribution. A cheaper canonical/admin override cannot masquerade as a provider price. Unsourced cheap layouts cannot hide a valid sourced layout. Filters run before the result cap. Source prices are recorded advertised layouts, not historical transaction medians, current quotations or live availability. Completion status is not inferred.

The last owned recommendation result preserves project UUID order and each project's search centre. Work corrections retain home, budget and bedrooms. Explicit new locations take priority; unresolved new place wording returns to normal cloud planning instead of searching the old office. First-two comparisons recheck publication and eligibility, preserve the referenced order, and never silently substitute another project. Mixed-centre references return to planning to establish a common comparison centre.

## Exact six-round cloud/browser replay

No response mocks or local generation model were used. The local runtime was configured for OpenAI `gpt-4.1`; the application does not persist provider-returned model IDs, so this is configured-model evidence, not a per-response identity claim.

| Round | Exact question | Observed result |
| --- | --- | --- |
| 1 | Help me find property locations that fit my daily routine. | Asked once for home and main destination. |
| 2 | I stay at Sunway Velocity and work in TRX. | Saved both roles; sourced Cochrane/TRX and Maluri/TRX connectivity, then asked budget. |
| 3 | Got MRT to TRX? | Reused the known endpoints; returned 1/2 Kajang Line stops without a train change; no generic criteria loop. |
| 4 | Purchase budget RM800,000, own stay, at least 2 bedrooms. Show me 3 suitable projects on the map. | Three real catalogue projects; same sourced two-bedroom layout prices in answer, cards and chart. Card focus selected the corresponding map property. |
| 5 | Actually I now work in KL Sentral, not TRX. Keep my home, budget and 2-bedroom requirement the same. Update the map and recommendations. | Work became KL Sentral; home, RM800,000 and minimum two bedrooms retained. Two eligible projects, not three invented matches. Explained the documented Muzium Negara pedestrian link and sourced 5/6 rail stops. |
| 6 | Compare your first two recommended projects for my routine and budget. Distinguish verified facts, estimates and current unit availability, and show your sources. | Compared Riveria City then Sunway Belfield without asking their names. Same order, UUIDs and prices in the answer/cards/chart; working map focus and source links. Distance estimates and unverified current availability remained explicit. |

| Round / project | Canonical UUID | Matching layout | Recorded price | Bedrooms / sqft |
| --- | --- | --- | ---: | --- |
| 4 — SKYLINE KL @ PUDU | `65d63b38-2fdc-4ce8-b8b9-6714c7b8a499` | `231bb403-6205-4d80-8364-b64a8b972b7a` | RM586,800 | 2 / 695 |
| 4 — Sunway Velocity 2 | `fe103b82-b2f2-46ca-89a6-9f72bdc6d519` | `59f72d7b-9f38-45b3-a2f4-108fb49c1ccf` | RM703,000 | 2 / 678 |
| 4 — Sunway Velocity 3 | `02e9875d-470c-46c1-ab18-e33996c8985f` | `afa3577c-dfd4-41be-b673-e824125e5163` | RM636,000 | 2 / 721 |
| 5/6 — Riveria City | `4e9539da-9606-45e8-854e-1627b15ced1c` | `5af861ea-9dab-4f32-bccc-5c4848567f3f` | RM604,400 | 2 / 486 |
| 5/6 — Sunway Belfield | `aeac3c58-8240-429b-948f-083173ad2b80` | `d4f664da-6a36-4a2b-95d4-c7e892aabce8` | RM775,000 | 2 / 797 |

The read-only server audit compared ordered `schemeIds`, project rows, chart-point `schemeId`s and numeric prices across rounds 4–6. It verified minimum bedrooms, purchase cap, source/layout identity, the changed work role and unchanged home. It also verified absence of Demo project names in those new results. Browser DOM values matched the source rows; the selected Sunway Belfield map popup showed RM972.4/sqft, matching RM775,000 / 797 sqft.

Property sources: [SKYLINE KL @ PUDU](https://propertysifu.my/newprojects/city-centre/condo/skyline-kl-at-pudu), [Sunway Velocity 2](https://propertysifu.my/newprojects/cheras/condo/sunway-velocity-2), [Sunway Velocity 3](https://propertysifu.my/newprojects/cheras/condo/sunway-velocity-3), [Riveria City](https://propertysifu.my/newprojects/kl-sentral/condo/riveria-city), [Sunway Belfield](https://propertysifu.my/newprojects/city-centre/condo/sunway-belfield). All prices above describe the recorded July snapshot, not a fresh fetch from these websites. KL Sentral access: [MRT Corp concourse map](https://www.mymrt.com.my/wp-content/uploads/2024/11/S15-MUZIUM-NEGARA.pdf).

## Additional regression and responsive checks

- Cloud/browser: `Can you show me 3 projects near Pasar Seni MRT station? Keep my budget and bedrooms.` searched **Pasar Seni**, not the saved KL Sentral office. It returned Sunway Belfield then Riveria City, retaining budget and bedrooms.
- An initial extra `Compare the first two, same filters.` preserved project order but incorrectly reverted its distance centre to the office. This was fixed by retaining the previous owned result's project-level centre. Replaying both extra questions then kept **Pasar Seni**, the new project order and the corresponding 0.73/1.51 km distances. The original failure remains in the local transcript; this is not a first-attempt/general-intelligence claim.
- Final assets were reloaded with the saved conversation. Desktop 1440×900 and narrow 390×844 had no page-width overflow. In narrow view, a project card opened Map with the correct name/PSF; Chat and the accessible price table remained usable. Temporary viewport overrides were reset.
- Regression tests cover source/canonical price conflicts, missing-provenance cheap layouts, draft/site suppression, ownership, rail-only intervening answers, project ordering, centre retention, stale saved areas, multiple centres, negative project requests, Chinese/English bedroom changes, catalogue cards with charts, neutral comparison headers, current-only map entities, type filters and earlier shortlist previews.

## Verification

- Related isolated PHP scope: **480 tests / 3,974 assertions passed** before the last search-centre metadata refinement. After that refinement and formatting, its affected catalogue/area scope passed again: **17 tests / 100 assertions**. The existing PHPUnit XML deprecation remains.
- Focused frontend scope: **31 tests passed** (Advisor and EvidencePanel).
- Production client and SSR builds passed from a non-Git temporary archive of the same frontend source, avoiding the existing source checkout's corrupt/iCloud Git metadata. Final Vue source hashes were checked against the built archive. Existing missing static-image references and unrelated unused-import warnings remain.
- Changed/new backend files pass Pint; the controller retains its pre-existing formatting outside this surgical change.
- Independent PHP and JavaScript reviews were performed; identified source, state, map-filter, asynchronous-preview and multi-centre issues were repaired. No production merge/deployment is included.

Full browser transcripts, read-only server payload checks and build/test logs are retained locally with this task. No new acceptance video is claimed.

![First-two comparison and map focus](evidence/catalogue-first-two-desktop-20261016.png)

![Narrow price table](evidence/catalogue-narrow-table-20261016.png)
