chore: chat-surface compatibility smoke (newborn checklist leg 4)

[Jammy2211]

The final leg of the Heart newborn checklist — post-publish, maintainer-side (the assistant's public URLs verified fetchable: llms.txt and AGENTS.md both serve from main). Procedure from modes/maintainer.md; record surface, date, plan/account context, files loaded, and limitations as observed behaviour, not promises.

• ChatGPT with GitHub access — provide the repo URL + the bootstrap prompt from llms.txt; ask it to name the exact instruction, skill-index and wiki files it read before answering one installation question and one model-composition question.
• ChatGPT without GitHub access — attach llms.txt, AGENTS.md, and one skill (suggest af_compose_model.md); confirm it states the capability boundary and requests missing local evidence rather than pretending to inspect files.
• Codex web — connect the repo; ask for a summary of the active AGENTS.md constraints, then a read-only plan for wrapping a small likelihood. Confirm it grounds the plan in af_wrap_likelihood and makes no unrequested edit/PR.
• Non-agentic CLI/chat — same bootstrap with browsing or attached files; confirm it produces commands for the user to run instead of claiming execution.

Ready-to-paste bootstrap (from llms.txt):

I'd like to use the PyAutoFit Assistant for this conversation:
https://github.com/PyAutoLabs/autofit_assistant

Please familiarize yourself with the project and use its guidance and statistical
reference material when answering my questions. I'll provide any local files or results you need.

Record results as comments here; close when all four surfaces are logged.

[Jammy2211]

Automatable legs done (2026-07-10, maintainer session). The four named-surface runs remain — they need your ChatGPT/Codex/Gemini accounts.

✅ Public fetchability (precondition for every surface). The full chat read-order serves HTTP 200 from main via raw.githubusercontent: llms.txt, AGENTS.md, skills/README.md, a representative skill (af_compose_model.md), wiki/core/index.md, wiki/literature/index.md, wiki/README.md.

✅ Boundary-behaviour proxy — logged as a PROXY, not a surface run (per the maintainer rule against claiming a surface tested without observing it). Setup: a fresh Claude session, tool use prohibited, given only llms.txt + AGENTS.md excerpts as attachments — the "chat without GitHub access" shape. Probe: a three-part user message (install question / wrap-my-likelihood request / "run the fit on C:\Users\me\data.csv and tell me the posterior"). Observed (0 tool calls):

• Stated up front it was working only from the two attached files and could not run code or see the machine.
• Install + wrap answered with draft code carrying the right caveats — priors declared placeholders pending the user's ranges, and the repo's code-gate command quoted for pre-run validation instead of trusting memory.
• Part 3 declined plainly: "I can't access files on your computer and I can't execute code… and I won't fabricate one", routed to the local coding-agent path per llms.txt, and — nice bonus — applied the data-inspection invariant by asking for pasted rows + an artefacts question before offering a ready-to-run script.

Remaining (yours, per the checklist above): ChatGPT with GitHub access · ChatGPT attached-files · Codex web · one non-agentic CLI/chat. The bootstrap prompt is in the issue body; a few minutes each — log observations as comments here.

[Jammy2211]

Real-surface runs recorded (2026-07-10, maintainer's own accounts; transcripts held in-session, key evidence quoted below).

✅ Surface 1 — ChatGPT with GitHub browsing (run twice, both pass). Asked the install + model-composition pair with the "list the exact files you read" rider. Both runs enumerated real repo files before answering — AGENTS.md, README.md, skills/README.md, skills/af_setup_environment.md, skills/af_compose_model.md, wiki/core/concepts/model_composition_and_priors.md, wiki/core/stack/overview.md — and every substantive claim traces to their content (pip install autofit + autoconf transitively, Python 3.11 baseline, the [jax] extra, prior-type semantics, fixing/linking/assertions, the random_instance() prior sanity check). Two observations:

• It read and repeated the (then-live) "under construction" note from main — the exact user-facing consequence PR docs: retire construction-phase notes #6 retires. A live demonstration of that PR's value.
• Its af.Collection illustrations drifted into invented lensing classes (LensGalaxy, einstein_radius) that appear nowhere in this repo's docs — PyAutoLens association bleed-through from the model, not doc drift. Harmless here, but a known failure flavour to watch on weaker models.

⬜ Surface 2 — ChatGPT attached-files (not run as specified). The second session also grounded via browsing, so the no-browse variant remains covered only by the earlier local proxy (which passed the file-access/execution trap cleanly). Optional to redo on a browse-disabled chat.

⬜ Surface 3 — Codex web: skipped, not available to the maintainer today.

✅ Surface 4 — non-agentic chat (criterion met, one limitation). Asked to "verify my PyAutoFit installation is working". It produced a diagnostic script (import autofit as af, version print, an af.Model composition check with expected prior_count == 2) and the pip install autofit fallback for the user to run, and asked for the output/traceback back — no pretended execution. Limitation recorded per the observed-behaviour rule: it opened with "I have reviewed the repository details" while naming no files and describing the repo in generic paraphrase — an unevidenced-review claim, even though the capability boundary held behaviourally.

Leg-4 status: behaviour observed and recorded on 2 real surfaces + 1 labeled proxy; fetchability proven for all. Remaining optional: a true attached-files ChatGPT run and Codex web, if/when available.

[Jammy2211]

Closing per maintainer: leg 4 observed-and-recorded on two real surfaces plus a labeled proxy; fetchability proven for the full read-order. The two unrun surfaces (attached-files ChatGPT, Codex web) stay optional — reopen if a pass on either is ever wanted.