feat(live): the placement interview, per-node prefetch and the rich bounded close (Lunaris Live, Phase 2c)

[PouyanJay]

Lunaris Live, Phase 2c: the placement interview, per-node material one node ahead, and the rich
bounded close. Phase 2b put the generative surfaces over an AG-UI transport; this makes a session
begin on a topic rather than on a compiled map, stay ahead of the learner, and end
deliberately: a recap, a mastery delta, a review schedule, and a next session that honours it.

Nine tasks, each test-first, reviewed by the project's agents and mutation-checked before commit.

T1	Walking skeleton: POST /api/live/sessions {topic} opens a placing session with the compile attached and the interviewer's first question; the web entry is session-first
T2	The interview: bounded, resumable, over both transports; warming and POST /{id}/advance (200 / 202); the compile that fails, and the one lost to another process
T3	Priors from the interview, credited then verified at the boundary before anything is built on them; the learner's profile briefs the tutor
T4	Per-node material prefetched one node ahead, kept per learner and map (live_materials), consumed once; the root prefetched while the learner is still being interviewed
T5	The close is a ceremony: the tutor's recap over the record (deterministic fallback), and the mastery delta from the beliefs stamped at open
T6	The spaced schedule: a review ladder written at close, shown to the learner, honoured by the next session's director; a finished map reopened before its day is a 409 naming the day
T7	The web ceremony over both surfaces: a real ending that names the review day, the interview labelled, the panel rebuilt when teaching begins
T8	Admission and metering: one reading of the ceiling for the turn and the prefetch, a topic-open that spends a compile slot and an opening or neither, the interview bound as a setting
T9	Every move and every close covered; the keyed eval extended to a real placement; and a claim credits the chain beneath it

What it does

• A session opens on a topic. The map compiles behind it while the interviewer asks what the learner already knows and wants; the mapper turns the answers into claims (never evidence) and a profile; the director credits a claimed chain to its boundary and checks the deepest claim with a graded retrieval before building on it. If the compile is slow the session warms and the surface advances it; if the compile fails the session closes and says so.
• Material is asked for one node ahead of the learner, under the session's own ceiling and keys, and kept per learner and map; a first turn on a concept whose material is there asks the tutor for words only. Later passes always generate fresh.
• The close carries the tutor's recap over what was covered, a meter that reads as movement (recall beside where the concept stood at open), and each concept's next review day. Every graded concept lands on a ladder (a day, then ×2.5 per rung, seven rungs); a session opened on or after that day retrieves it first. Both surfaces say "Come back on Thursday 20 August for Prior and Update."
• No new cost subject: interview, mapper, prefetch and recap are the session's money under LIVE_SESSION and the $2 ceiling binds them all, on both transports.

The decisions worth arguing with

• A claim credits the chain beneath it, and is checked before anything is built on it (AD27). The keyed eval's first run put claims on the gradient, the chain rule and backpropagation and none on "slope and derivative", which nobody had asked about, and the director introduced the root. Now a credible claim credits what it stands on, the check is the claim that implies it, and a demonstration credits downward the same way (a MET on top of a chain is evidence for the chain). A hesitant claim still credits nothing.
• The schedule is a ladder, not a forgetting curve. Beliefs still do not decay across sessions; the due date is the only cross-session pull. Days are named in UTC on every surface; the learner's timezone is a later item.
• A review answered mid-session moves to tomorrow, not to nothing, so a session that never closes cannot drop a concept off the ladder.
• The panel is rebuilt across the placing → teaching boundary: its thread opens on the turn it mounted with, and the first lesson arrives through the host's poll.

The eval ran

test_a_fluent_learner_is_placed_past_the_root and test_two_learners_on_one_concept_get_material_in_different_bands, real interviewer, mapper and tutor: the fluent learner is interviewed, placed with 13 to 19 claims and a profile, and opens on a retrieval at the boundary of what they claimed; the same concept is composed in different Tier 2 bands for the placed and the fresh learner, with different material. Digest in .eval/live-session/placement-fluent.json.

Gates

Python 1407 (packages/live + apps/api) · tests/db 97 on real Postgres · web 1696 · ruff, format, tsc, eslint, Prettier, supabase db lint clean · every task mutation-checked (T5 15, T6 26, T7 11, T8 11, T9 6, all killed after the review rounds).

Migrations (expand-only, in order)

live_sessions.status widened with placing and warming; live_knowledge.prior; live_materials (RLS select-own); live_knowledge.review_stage (CHECK 0..7) and due_at.

Rollout, and the Phase 2 exit

Merging deploys to dev. The Phase 2 exit criterion that is the operator's and the learner's, not the suite's: LIVE_RUNTIME_ENABLED=true on dev, VITE_COPILOT_URL set to the printed runtime URL, the SPA rebuilt; then one real browser session on /live?topic=…, interviewed, taught, closed with a day to come back, and rated. New dials are documented in documentation/deployment.md.