M6 acceptance: onboard 3–5 unaided Python learners

[udhawan97]

Goal

Collect honest first-run evidence from 3–5 learners who built a small Python project with AI. This is the final Phase 0 acceptance item; Phase 1 must not be promoted before it passes.

Start here

• latest tester release — collect evidence on v0.7.0 or later (see the version note below)
• install with uvx codemble; no Git tag or Node.js needed since v0.3.0
• ten-minute tester guide
• use the Early-tester report issue form; never share private source or keys

Acceptance evidence

• Tester 1 lights a system unaided and reports first confusion verbatim
• Tester 2 lights a system unaided and reports first confusion verbatim
• Tester 3 lights a system unaided and reports first confusion verbatim
• Tester 4 result, if recruited
• Tester 5 result, if recruited
• File actionable confusion as linked issues without paraphrasing away the learner's words
• Update CLAUDE.md only after at least three independent unaided successes

Version floor

Evidence must come from v0.7.0 or later. Earlier tags are disqualified for reasons that bear directly on the step this issue asks testers to perform:

• before v0.2.1 — multi-answer checks offered no wrong option, so select-all lit a region without proving understanding
• before v0.6.4 — a missed check printed the parser's own answer, and the same question then accepted it
• before v0.7.0 — the map's language stripe rendered as the box background, the open Key covered the zoom controls, Escape did not step back on the Map, and Fit landed at an unreadable 7% on a tall project

Guardrail

An install, parser, explanation, Lens, citation, or check correctness failure is not “tester error.” Treat it as product evidence and file the smallest reproducible issue that does not expose private code.

[udhawan97]

Reopening: closed in error. No unaided learner runs have been recorded — all seven acceptance boxes are still unchecked and no tester reports were filed. Technical completion of Phase 1 does not satisfy this criterion.

Note for anyone picking this up: install evidence should be collected on v0.2.1 or later, not v0.2.0. The v0.2.0 tag predates the fix for multi-answer checks that offered no wrong option, which let select-all light a region without proving understanding — exactly the step this acceptance issue asks testers to perform.

[udhawan97]

Refreshed the tester path; the acceptance criterion itself is untouched and still unmet.

No unaided learner runs have been recorded. All seven boxes remain unchecked, no tester reports have been filed, and CLAUDE.md still states the milestone has not advanced. Nothing below is evidence.

What was blocking a tester who arrived today:

• The issue pointed at the v0.1.0 release, six releases behind, and at a v0.1.0 early-tester report form.
• The report form (.github/ISSUE_TEMPLATE/early_tester.yml) was titled v0.2.0 and its required Install route field offered only "pipx from v0.2.0 Git tag" and "uvx from v0.2.0 Git tag". Codemble has published to PyPI since v0.3.0, so both were dead paths — the one form whose purpose is collecting honest install evidence forced a false answer.

Fixed in 599b333: the form is version-agnostic, lists the real install routes, and now requires the version tested (matching bug_report.yml) plus the project language, since JS/TS and mixed projects ship now.

The body also gains an explicit version floor of v0.7.0, with the reasons stated: pre-0.2.1 let select-all light a region without proving understanding, pre-0.6.4 printed the parser's answer on a miss and then accepted it, and pre-0.7.0 shipped an invisible language stripe, a Key that covered the zoom controls, and a Map where Escape did not step back. Each of those distorts precisely the step this issue asks testers to perform, so evidence from those tags would not be worth what it looks like.

Recruiting real learners is the remaining work, and it is not something the tooling can produce.