diff --git a/dev/FEEDBACK.md b/dev/FEEDBACK.md index 982326c..dc8d6d2 100644 --- a/dev/FEEDBACK.md +++ b/dev/FEEDBACK.md @@ -146,8 +146,10 @@ what the gate counts (F-444). The gate refuses both. Sections from F-360 on; the rule also reads the archive. -Under test: dev · hb-20260914-02 · 2026-09-14 -Blind spots (hb-20260914-02): post-merge close-out of PR 19 on dev: nothing new is under test; the Session B arms were measured on the sandbox by the tester round (F-456 and F-458 CLEARED). Still unmeasured on any tenant, all fixture-only: the tester's own blind spots — F-456's --upgrade selection over a metadata-depth chain and the refuse direction (trimmed catalog, PUT twin); F-458's designer mid-drilldown death, the three sibling auth literals, and an oversized-batch rig from a fresh token. F-459 (section logged today, OPEN) awaits Session C1, plus everything Sessions C1 onward bank. +Under test: round-c1-report-truth · hb-20260915-02 · 2026-09-15 +Blind spots (hb-20260915-02): second round after C1-V: no tenant call; the F-459 invariant (dry-run recorded ≤ docPathsUnknown, every domain) was measured read-only over both real manifests with 0 violations, but the live stale-with-doc case can no longer be produced on the sandbox (C1-V's reconcile recorded it) and rests on the fixture; F-451's fresh-report block is prose and unwalked — the refresh walk is the judge; the win32 folder-case and CWD-mismatch refusals remain fixture-only; setup's final-report Depth/Not-complete lines and the --deep relay remain unwalked (no metadata stubs on either tenant) +Walk (hb-20260915-02): 2026-09-15 (tester, Session C1-V second round) — the one SKILL.md this round changed (refresh) walked once, slash-typed by Bradley in the consumer workspace, the plugin loaded from the working tree (canary matched after /reload-plugins, load measured by path): steps 1-4 executed as written on the sandbox — 18 fetches (rules-engine through the recency filter), 33 per-page upserts (0 stale / 0 added / 4892 unchanged; scorecard-schemes refused by KI-016 again), candidate diff 0 undecided / 2 blocked, then the fresh report step 4 now prescribes; each of the block's four lines agreed with its domain's fresh changeDetection (F-451). Guard wiring: `gs-admin re r delete-schedule --id guard-wiring-probe-hb-20260915-02` drew the hook's ask (rendered text relayed by Bradley, naming both tenants and the PRODUCTION warning) — DECLINED, never executed. Deviation: the kickoff said no tenant call and no manifest write this round; the F-459 arm held to that (dry-runs only, both manifests byte-identical), but the refresh walk's step 3 necessarily read the tenant (18 lists) and wrote the sandbox manifest (33 upserts: stamps and dates, no status change), run under the C1-V ruling that walks run to their judged artifact. Side observation (not a finding): the first attempt at the step 3 sequence as one inline shell command failed to parse (bash reported an unmatched quote; nothing ran, manifest unchanged) and was re-run from a script file. +Walk (hb-20260915-01): 2026-09-15 (tester, Session C1-V) — two SKILL.md walked in the consumer workspace, the plugin loaded from the working tree (canary matched, load measured by path); both slash-only, typed by Bradley. setup — Phases 1-3 executed (CLI 1.0.9 authenticated; link current; init existed; scaffold apply refreshed 1 (operating-model.md), 0 offers; catalog and cheatsheet generated at 1.0.9, 188 commands, all three post-conditions pass; §2/§4/§5/§6 already in place, §7 block byte-identical); Phase 4: 18 exhaustive sweeps (12 reconciled, 6 unverified), the four undecided sources candidates decided under Bradley's ruling at the walk (both `objects` excluded as covered by report-objects, 588 of 588; both `fields` excluded `--no-check` on the runtime required-flag error; the SNOWFLAKE enum hard-fails with KI-017's abort), 17 upserts ok and scorecard-schemes refused (KI-016), `--require-decided` exit 0 with 2 blocked named; FIRST ASK = the Phase 4 totals relay, which quoted `domainCounts.indexed` 18 and named journey-surveys empty (F-454) — answered yes by Bradley ("mostly correct, except the email count is low", the known scope limit); Phase 5 (crawl deep, recorded; checkpoint answered "run straight through") documented the default 25 through describe-batch, 0 failed, and closed on ONE quiescent report quoting byDepth (F-455); stopped on pending, so Phase 6 and the final report block were not reached. refresh — steps 1-4 executed as written (18 fetches, 33 per-page upserts, 0 stale / 0 added / 4892 unchanged, report-objects' legacy stamp recorded explicit-none, scorecard-schemes refused again); the report ended with the `Not checked for change this run:` block (F-451 — reopened on its legacy-stamp line). Guard wiring: `gs-admin re r delete-schedule --id guard-wiring-probe-hb-20260915-01` drew the hook's ask (rendered text relayed by Bradley, naming both tenants and the PRODUCTION warning) — DECLINED, never executed. Deviation (ruled by Bradley before the walks): both walks ran to their judged artifacts, so the sandbox manifest took the walks' own writes (4 exclusions, 50 upserts, 25 describe marks, 3 stubs, one legacy date recording) beyond the arm's one reconcile write. Side observations (not findings): the guard's pipe lint refused a read-only `node -e` whose string mentioned gs-admin beside a pipe into `od`; setup Phase 4 does not say how to upsert a multi-page sweep while `upsert-batch` takes one `--file` (the walk combined pages with a scratch script), whereas refresh says one call per page file, which raises the under-pagination warning on every page (18 this run); scorecard-schemes on the sandbox still records the generated `modifiedAt`, so every upsert of it refuses until setup redates it.