Revisions batch: on-ramp + model currency + D13 scope row (spec), six guide rules, first full live-AC run (6/6 after the thread-identity fix)

New course: Building On-Device AI Memory with Qdrant Edge (5 lessons,
Dylan Couzon). Materials are the notebook dump (L3-L5 + helper.py +
requirements.txt + ro_shared_data), the five transcripts, the slide
descriptions, and environment.md — the AI Coding Lab runtime the build
targets.

The spec is filed as the variant, spec.coding-agent-lab.md, because it is
generated against that environment rather than for an unknown one. Two
consequences:

- The Decision Ledger carries four rows, not ten. Five of the six
  learner-context dimensions are fixed by the lab (project and goal are
  §1, data is the §5 fixture corpus, the runtime is the container, and
  the course uses no LLM at all, so there is no provider to pick). What
  remains is course decisions: on-device vs cloud sync, one store vs a
  store per purpose, similarity vs similarity+freshness, plus the
  scope-boundary row that presents §1 at the gate.
- Environment facts are labelled [environment] throughout and given
  their own CTX-E provenance row, so the course is never credited with
  port-4000 rules, the container's provider keys, or the choices-fence
  question mechanism.

The slides earned two changes the other materials could not: R7 now
states that the filter runs inside the query rather than trimming its
results, with AC7b to discriminate the two (a garden-filtered food query
must still return the lone garden note); and R5 records that the course
drafted a merged single-inbox display and dropped it — it appears in no
video, while the shipped code renders separate lanes and names the
blended list as the thing not to do. The decks also carry a third,
abandoned lesson numbering, noted in CTX-E so it is never cited.

No canonical spec.md yet, so /prepare-build will refuse on this course
until one exists.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The Qdrant course's lab-targeted spec was not reproducible. The transcript
of the session that produced it shows the skill read the base guide plus
materials/notebooks/ and materials/transcripts/ and nothing else —
environment.md was never opened by any tool; it was in the model's context
only because the user's message carried it. Every lab-specific decision
(dropping five of six learner-context rows, inventing an [environment]
label, naming the choices fence as the gate's tool, ignoring the
lab-packaging half of environment.md) was improvised mid-generation and
recorded nowhere. The spec's header still claims conformance to guide
8e44ecb, which mandates six learner rows; it has one.

This makes that mode real, without touching the base guide.

- environments/coding-agent-lab.md: environment.md moved out of the course's
  materials/. One environment serves many courses, and materials/ is what
  /eval-materials-vs-build scores a build against, so a runtime doc there
  would eventually be scored as course fidelity.

- references/spec-generation-guide.coding-agent-lab.md: the overlay. It
  overrides exactly two base passages — §3's input gate and §5.5 step (f) —
  and quotes each verbatim so staleness is mechanically detectable. Its
  precedence rules say the base guide wins everywhere it is silent, and that
  an unanticipated conflict is reported as an overlay gap rather than
  resolved silently at generation time. Everything else it contributes is
  additive: the Environment Resolutions table, an [environment] provenance
  label, §14 checklist items, and the runtime facts distilled from the
  source doc's §4, §5.1, §6 and §9. The choices fence is an input to base
  §6.0 ("use a structured tool if you have one"), not an override of it.

  The overrides are two-phase where it matters: all six learner-context rows
  are still derived in full, then resolved as ANSWERED / CONSTRAINED / KEPT,
  with an ANSWERED row's invariant required to land somewhere named. Ordering
  is load-bearing — the runtime facts are not to be read until derivation is
  done, because a dimension never written down cannot be resolved, and that
  is exactly how five rows vanished untraced the first time.

- Both SKILL.md files take --env=<slug>; the no-flag path is unchanged word
  for word. prepare-build normalises the filename when it snapshots, so the
  build agent and both eval skills never learn a variant was involved, and
  its announcement string is untouched because extract.py regex-matches it.

- .gitignore's negation becomes a wildcard, so a new slug needs no edit.
  CLAUDE.md, its tracked twin AGENTS.md, and README.md document the flag.

Verified: zero diff to the base guide; both overlay quotes match it
verbatim; the extract.py bookend still matches; no existing SKILL.md line
reworded. Regenerating both Qdrant specs is the next commit and the real
test of the overlay.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The course had only the lab variant, so there was nothing to compare it
against and no way to answer "what does the lab actually change" except by
reading prose. It also could not be built at all: /prepare-build needs a
spec.md.

spec.md is generated agnostically from the same materials under guide
8e44ecb: nine Ledger rows — the six learner-context dimensions (project,
data/inputs, goal, model/provider, environment, scope-boundary) plus the
same three course-owned rows the variant carries (memory locality,
topology, ranking). The variant has four. The five-row difference is the
environment's whole effect, and it is now a diff between two files rather
than a claim.

The variant is brought into conformance with overlay b0d88d6: its
hand-written "Four rows, not ten" note becomes the Environment Resolutions
table the overlay mandates, accounting for all six dimensions as ANSWERED,
CONSTRAINED or KEPT with each invariant's new home named. Its header now
states honestly that the file predates the overlay, that its lab decisions
were made by hand rather than by rule, and that a clean-room regeneration
is still pending.

Also fixed in both, from auditing them side by side:
- AC8b — a typed description must retrieve a photo. R4 claims a question is
  embedded once per space, but only the image-to-image path was tested; the
  cross-modal half, which is the course's most distinctive capability, had
  no criterion in either spec.
- AC18 was orphaned in spec.md (no rule cited it); R2 now does.

Note the count: nine rows agnostic. Earlier in this work that number was
asserted from arithmetic (six plus three) before any agnostic spec existed.
It happens to be right, but it was a guess then and is an artifact now.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

First real test of --env=coding-agent-lab: a subagent with no memory of the
authoring session generated the spec from the guide and overlay alone. It
worked — and found three defects in the overlay, two of which it could only
flag rather than obey.

1. The overlay contradicted itself on reading order. OVERRIDE 2 said "do not
   read the RUNTIME FACTS section until Phase 1 is complete", but the same
   file's header said read it in full and the mandatory staleness check forced
   reading it before generating. Unsatisfiable: you cannot un-read part of a
   file you were told to read, and cat/Read return it whole.

   RUNTIME FACTS now lives in spec-generation-guide.coding-agent-lab.runtime.md,
   opened at OVERRIDE 2 Phase 2. The ordering is enforceable because it is
   physical — the file simply isn't open yet. SKILL.md resolves both halves at
   step 0 and reads only the rules half at step 2.

2. model/provider resolved wrongly, and this is the substantive one. ADD 1
   hardcoded CONSTRAINED. But this environment fixes the model the coding agent
   runs on and supplies ambient LLM credentials; it fixes nothing about models
   the course's own pipeline uses. For a course whose models are local encoders
   that deletes a real learner decision — one with a real invariant and a real
   switching cost, since changing an encoder means re-embedding everything.

   The cell is now conditional: app needs an LLM -> CONSTRAINED; course pipeline
   needs none -> KEPT as an ordinary Ledger row, plus a rule that the ambient
   keys must not be read. The runtime companion's own section now defers to that
   test instead of implying the CONSTRAINED answer.

   The clean-room run complied with the wrong table and reported the gap, which
   is precedence rule 3 working as designed — it just produced a spec missing a
   legitimate row.

3. Base §14's learner-row checklist items had no rows to apply to once five
   dimensions were resolved. ADD 3 now re-points them at the Environment
   Resolutions table: a resolved dimension discharges the obligation at its
   named home, a KEPT one as an ordinary row. Neither line is waived.

Base guide: still zero diff. Both OVERRIDE quotes still match it verbatim.

Not fixed here, filed instead: four base-guide findings (row-grouping
ambiguity, no category for a fixture-inexpressible input, §5 vs §14 on version
pinning, and the surfacing-bar divergence itself) and the slides glob. Fix and
test travel together; those need a promotion pass, not a hand edit.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Both files in this course were written by hand and audited by their own
author. Neither had ever been produced by running the skill. Two subagents
with no memory of the authoring session generated them from the guide,
the overlay and the materials alone, forbidden from reading the prior
outputs. Their work replaces mine.

  spec.md                   738 hand-written ->  590 generated (11 Ledger rows)
  spec.coding-agent-lab.md  789 hand-written ->  805 generated  (6 Ledger rows)

The three overlay fixes in 58ecfb9 all verified against the run that could
not have known about them:

- model/provider now resolves KEPT, and the variant carries a real
  "Encoder set" Ledger row (D1) whose Options mark the GPU-class and
  hosted-API realizations impractical-in-this-environment rather than
  deleting them. Under the old overlay this row did not exist at all.
- The .runtime.md companion was opened after all six dimensions were
  derived, and the run reported that ordering unprompted. The reading-order
  contradiction is gone.
- No orphaned §14 checklist items reported.

Both specs audited structurally before adoption: JSON Schema blocks parse,
every AC traces to a business rule (the agnostic spec added an explicit
rule column to its AC table, which is mechanically checkable in a way my
prose citations were not), §0 gate present and first, CTX pack complete,
four-way failure encoding mapped.

The two runs also disagree with each other, which is the finding. On the
same course and the same base guide: the agnostic run made the memory store
a `realization` row and the id scheme a `contradicted` row; the variant run
produced zero `realization` rows, reasoning that a keyless pip install and
model download does not clear §3's heavy bar. Both baked freshness ranking
in as a business rule where I had made it a Ledger row.

Findings filed, not fixed: #28 (four base-guide defects, including the
surfacing-bar divergence) and #29 (the slides glob). Six further overlay
gaps surfaced by these runs — the deliverable's shape, ADD 2 forbidding a
co-label the course and environment both require, ADD 3's empty-invariant
branch, the "Standalone Takeaway" title in lab mode, the overlay's missing
self-commit, and whether environments/coding-agent-lab.md may be opened —
are not yet filed.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

…device-ai-memory-with-qdrant-edge

Add the Qdrant Edge course, the --env harness, and both generated specs