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)

[Carr1005]

The revisions batch from the tester-feedback triage (#24) plus the promoted #18 findings — spec edits and their guide generalizations, validated Option-B style (surgical edits + clean-room Opus regens; no adoption, no rebuild). Also includes the lab's first full keyed live-AC run, which caught and fixed a real design gap.

Spec (3 commits)

• 2835655 — on-ramp + currency: the opening blockquote gains the orientation lead ("what you're building" as a user-visible scene) and the fastest-path/completion sentence (offline suite green = the recommended baseline is complete; keys unlock only live tests/live use); inline heavy-setup notes on D11 Oracle and D12 Tavily; D4 Options name providers, never "current" model names ("expect the build agent to recommend current ones at the gate"); CTX-D gains the request-parameter-compatibility perishable (the temperature-rejection class); fixtures/llm-stubs.py → llm_stubs.py.
• 809dea8 — D13 summary-memory-scope row (owner decision, first contradicted-category row): thread-scoped default (the app path) vs the Lesson-5 notebook's unscoped global pool — both course-demonstrated; stakes test passes because the choice changes what a read returns. Side edits: contradicted-group paragraph, R18 defers to D13, §5 coverage note, CTX-C11 cross-ref.
• R11 harness-known-identity sentence — added mid-batch after the live AC21 run caught the gap (below), per the spec's own living-document rule.

Guide (1 commit, 0cd5ff2)

Six course-agnostic rules: G1 orientation-block template + reader test; G2 heavy options state setup weight inline at choice time; G3 model currency is the generation-vs-gate timing split, CTX-D parameter-constraint category; G4 identifier formats are Pass-A parameters, alphabet included; G5 fixture module filenames must be importable identifiers; G6 the contradicted route's worked storage-scoping example. Reference-style §14 lines.

Validation

• 3/3 clean-room Opus regens: all six new V-properties present in every regen (B full-scored 30/30 PRESENT, 0 PARTIAL with quoted evidence + orchestrator spot-audits; A and C grep-verified). G4 closed the id-alphabet cross-model gap 3/3 (was 0/3); G6 held the scope row 3/3; regen-B is the first exact-canon 13/13 Ledger in any pass. Carried residual expressed as documented (A +1 row, C −1) — logged, not fix-rounded.
• Stage-1 gate probe (Opus): cold-start build agent still stops at the gate with the new blockquote; relayed the completion criterion correctly.
• Matrix: docs/research/guide-revisions-batch-matrix.md (decision log incl. F8 parked and Recording: Claude Code transcript compaction can erase the /prepare-build announcement bookend #19 excluded by owner).

Live-AC run (run-09, owner-supplied key) — the headline

First keyed run: AC4, AC9, AC19, AC22, AC24 passed live. AC21 — the course's cross-session continuity demo — failed deterministically: the assembled context never tells the model its thread id, so it called summarize_and_store with a guessed "current_thread" and the tool honestly consolidated nothing. The oracle worked exactly as designed: the spec gained the missing constraint (R11: harness-known identities are injected, never model-supplied), the run-09 build was conformed (~20 lines), offline suite re-green 21/21, and AC21 passed on rerun — live suite 6/6. The course-agnostic guide rule ("harness-known identities are injected, never model-guessed") is recorded in the matrix as the next pass's first candidate, per fix-and-test-travel-together.

Issue routing

• Tester-feedback triage: converged statuses, owner decisions, and the batch slate (revisions + next pass + measurement) #24: batch items A1–A4 delivered here; C-watchlist unchanged; live-AC item closed (6/6).
• Guide residuals from guide-pins-v2: route closure, constants provenance, fixture churn, gate-semantics reform #18: id-format (G4) and fixture-filename (G5) findings implemented; summary-scoping canon admission implemented (D13 + G6 + P1→13); F8 remains parked by owner decision.

🤖 Generated with Claude Code

[Carr1005]

Two fixes added mid-review (commits eb953a0, 0813752), both prompted by a disposable-clone trial run after this PR opened: the spec was copied alone into an empty folder and built by a fresh session with zero project context. That build came out green (50 passed / 6 skipped) and scored 9/10 on an independent eval — and it exposed two defects in the assets this PR touches.

1. eb953a0 — R11's harness-identity rule was too narrow (my defect from this PR's own R11 commit). The wording named summarize_and_store inside the normative clause. The trial built the general mechanism — injected params stripped from the model-facing schema, harness values overriding model-supplied ones — but applied it to exactly that one named tool, leaving expand_summary unscoped and quietly weakening D13's thread isolation. That is wording-discipline violation #2 (anchored to a course specific rather than a concept). The rule now covers every thread-scoped agent-triggered operation, requires the parameter be absent from the model-facing schema, and states explicitly that identifiers the model legitimately holds (a summary id read from a [Summary ID: ...] reference) stay model-supplied.

2. 0813752 — the gate checklist is now a durable artifact. §0 told the build agent to print the resolved-decision checklist; §7 then cited "the printed resolved-decision checklist from §0" as evidence — demanding durable proof of something that only ever existed on screen. The drift this caused is measurable: across three builds the checklist landed in resolved-decisions.md (a probe build, spontaneously), inside BUILD-REPORT.md (run-09), and partially inside README.md (the trial — 3 of 13 rows) — three places, two of them files no spec mentions, so no eval could ever verify the gate ran. It is now printed and written to a fixed-name file, resolved-decisions.md, in the build folder, with the reasoning in the guide's §6.0 rule. Owner constraint captured in the wording: the file is a record, never an input — a new or resumed build re-runs the gate from the spec and never reads a previous run's checklist, and each gate run overwrites it (no accumulation; build folders are gitignored anyway).

Files to re-read: courses/agent-memory-building-memory-aware-agents/spec.md (§0 step 5, §7 evidence line, R11) and .claude/skills/generate-spec/references/spec-generation-guide.md (§6.0 checklist-echo rule, the emitted §0 template step 5, the §0 anatomy comment, the express-lane bullet, two §14 audit lines).

A template-conformance validation regen is running against the new guide commit; its result and the accepted residual (the behavioral check — does a build actually write the file — rides the next trial rather than a paid probe build) are recorded in docs/research/guide-revisions-batch-matrix.md. Four further trial findings are deferred to the next pass and logged there, not applied here.

🤖 Generated with Claude Code