Replace placement questionnaires with a compact interviewer conversation

[rosshd]

Parent: #58
Depends on: #68.
Related: #59, #65.

Problem

The current placement opens with setup and calibration, then requires separate clarification and plan fields before the learner reaches the editor.
This feels like a questionnaire rather than an interview, while the required prose is classified with brittle keyword, punctuation, and length heuristics.

Removing communication entirely would also be wrong.
Clarifying ambiguity, stating an approach, and responding to an interviewer are explicit interview skills in the canonical skill graph.

Goal

Implement a compact 10-12 minute conversational placement sample that preserves authentic clarification and route explanation without reproducing the fixed seven-stage wizard.

Learner journey

1. Show one concise introduction with expected duration, the sequence talk briefly -> code -> debrief if needed, and the stop command.
2. Present the original or licensed problem in an interviewer-style exchange.
3. Let the learner ask bounded clarification questions and receive contract-consistent answers.
4. Ask for one concise approach summary before coding.
5. Transition directly to the configured interview editor.
6. Record communication evidence separately from correctness evidence.

Technical scope

• Add the v2 conversation state and CLI routing over Introduce a versioned compact-placement lifecycle with v1 resume compatibility #68.
• Remove mandatory self-report calibration from the critical path.
• Support multiple clarification turns without creating graded-answer state.
• Preserve a deterministic offline interviewer for problem-contract answers.
• A configured model may improve conversational phrasing, but cannot change the authoritative problem contract, lifecycle state, or scoring outcome.
• Use explicit commands such as /code, /stop, and /discard without front-loading every recovery command in the opening banner.
• Record the learner's final approach summary as bounded attempt evidence.
• Keep communication evidence independent from implementation and mastery evidence.
• Save every learner turn before provider calls or editor launch.

Likely surfaces:

• src/openlearn/cli.py
• src/openlearn/interview_prep.py
• src/openlearn/interview_attempts.py
• tests/test_cli.py
• tests/test_interview_prep.py
• tests/workflows/test_interview_journey.py
• manual-tests/interview-placement.md

Acceptance criteria

• A new placement states the expected 10-12 minute shape before it starts.
• The learner reaches the problem conversation without a mandatory self-report calibration answer.
• Clarification supports more than one bounded turn and never contaminates pending tutor-answer state.
• Contract answers remain correct offline and cannot be altered by model prose.
• The learner provides one concise route or plan summary, then enters the editor without another questionnaire stage.
• Blank input, /code, /stop, /discard, EOF, and interruption have clear durable behavior.
• Communication evidence is recorded separately and does not grant mastery.
• Provider failure preserves the learner turn and falls back to deterministic interviewer behavior.
• Existing non-interview courses and normal tutor turns receive no placement prompts.
• PTTY tests cover happy path, multiple clarification turns, offline fallback, stop/resume, provider failure, and cancellation.
• make check passes.

Out of scope

• Audio capture or speech scoring.
• Autocomplete enforcement inside the editor.
• Full timed mock interview behavior.
• Adaptive debrief selection and result-card presentation.
• Company-specific interviewer simulation.

[rosshd]

Implementation dependency discovered while building #68:

coding_activities.py currently validates the historical placement evidence stage names and rejects v2 conversation before activity evidence can synchronize into the profile.

#68 intentionally leaves production placement on v1 and does not broaden into this shared activity contract. When #69 flips the CLI to v2, it must:

• extend the coding-activity interview observation contract to the versioned placement stage set;
• bind accepted stage names to the activity/placement lifecycle version rather than one global allow-list;
• reject cross-version stage injection;
• cover activity publication, restart reconciliation, duplicate replay, and terminal-attempt behavior for v2 conversation evidence.

This is a required #69 acceptance seam, not an optional cleanup.

[rosshd]

Manual test evidence - 2026-08-04 placement run

Tracked findings: MT-001, MT-002, MT-006.

Observed on current local main:

• Course creation still asks goal, role family, level, date, weekly minutes, session minutes, and confirmation before placement, even though most values have safe defaults and advanced details are described as editable later.
• The placement then adds a mandatory free-form calibration before clarification and planning.
• A dictated or split answer is consumed one line at a time. The second part of the learner's clarification was accepted as the plan, and the second part of the plan arrived at implementation> and was rejected. The stage advanced on partial text without an explicit finish action.
• At implementation, any non-command text only repeats Press Enter to open your configured editor..., producing a loop when the learner naturally says that coding is complete.

Expected:

1. The default path asks only choices needed to select a safe starting route; advanced profile fields remain editable later.
2. Clarification and route discussion support natural multi-turn or split input without silently assigning text to the next stage.
3. The transition into coding states the editor return/submission behavior before launch and offers one clear action after a failed run.
4. A learner saying done, implemented, or similar receives actionable guidance instead of an identical prompt loop.

Regression coverage should replay the exact split-input sequence from this run in a PTTY test.

Classification: defect plus interaction polish. Severity: high because communication evidence is attributed to the wrong stage and the learner cannot discover how to progress.

[rosshd]

Superseded by the user-approved short reasoning placement now implemented as reasoning-placement-v3. The current placement intentionally takes about five minutes, uses clarification plus a concise route summary, and does not open an editor or score coding fluency. Real coding remains in later course practice, so the 10-12 minute talk-to-code placement contract in this issue is no longer the intended product behavior. The current replacement is covered by manual-tests/interview-placement.md and tests/workflows/test_interview_journey.py.