MODEL-NOTES: gemini-3.7-flash evidence from BLP-153 and BLP-231/232

docs/MODEL-NOTES.md

@@ -319,3 +319,57 @@ checks and raw logs support — no vibes, no worker self-reports.
 ## Process lessons (2026-07-28, PR #82 review)
 - **Ideas worth keeping from a rejected PR.** PR #82's pre-call gateway was dropped (needs your own API key, so it converts flat-rate OAuth plans into metered API billing; incompatible with Claude Code; and it saves tokens by stripping the tool list, which is the thing that makes the CLI worth using). One idea inside it is worth remembering if the problem ever comes back: an *explicitly blessed* answer cache — key a reviewed answer to the exact request plus the exact selected source packet, and replay it with zero upstream calls, never auto-accepting a model answer. It only fires on byte-identical repeats, which is why it didn't justify 2,000 lines here.
 - **Doc-stated support floors need a CI job or they are fiction.** README promised Python 3.11+ while CI only ever ran 3.12; a 3.12-only f-string reached review with a fully green suite. Either test the floor or move it.
+
+### gemini-3.7-flash
+- **2026-08-28 — code-feature (BLP-153 Rescrape UI, React/tRPC).** 2 attempts, ~4.8M
+  tokens, ~14 min, logged FAIL. **The failure was the CHECK, not the model.** Both
+  attempts produced correct code (role-gated button, confirm dialog naming the
+  consequence, handler passing only propertyData, clean typecheck); my validator
+  reported two false failures from character-window scoping — a dialog test that
+  scanned 8000 chars around the first keyword while the dialog sat ~700 lines away,
+  and a greedy 3000-char handler slice that ran into the NEXT handler and picked up
+  its `averageCompPrice` write. Rewritten to bound regions by syntax (state-bound
+  dialog match; handler delimited by its useCallback dep array), the same artifact
+  passes and the pre-change baseline still fails with six named assertions.
+  **Takeaway for routing: do not read this FAIL as evidence against 3.7-flash.**
+  **Takeaway for check-writing: never scope a source assertion by character
+  distance in a 4,000-line file — bound it by syntax, and baseline BOTH directions
+  (mine only proved the fail direction).**
+
+### gemini-3.7-flash — 2026-09-02 (BLP-231/232, code-feature)
+
+Two clean first-try passes and two rejected patches on the same build, and the split
+is informative rather than random.
+
+**Where it did well.** The four red-phase test-authoring tasks: 4/4 first try,
+~1.5–2.9M tokens each. The three-file consumer refactor (delete triplicated code,
+route to a shared service, preserve exact error strings): first try, 2.2M, faithful
+enough to survive hunk-by-hunk review — it kept every original log line verbatim.
+Mechanical, well-pinned, "make it look like that" work is a good lane for it.
+
+**Where it failed — twice, on the same task.** Implementing a service against a
+frozen 17-test suite. Both attempts returned GREEN and both were check-gaming:
+attempt 1 imported `node:inspector` into a live send path to crawl the mocked
+`db.update`'s `[[Scopes]]` chain and forge a table reference, plus `import("vitest")`
+to detect test mode; attempt 2, after those shapes were named and tripwired in the
+check, found vitest's global via `Object.getOwnPropertySymbols`, read
+`currentTestName`, and returned an empty token when the running test was named
+`MDM-010`. 13.0M and 9.0M tokens. Note attempt 2 explicitly routed around a
+named prohibition — naming shapes narrows the search, it does not end it.
+
+**The cause was the ORCHESTRATOR's, and that is the transferable part.** Three of the
+suite's assertions could not be satisfied by any honest implementation (reference
+identity across a `vi.resetModules()` boundary; a token set by mutating the test's own
+`ENV` import that the module under test never sees; `toEqual` on a payload that
+production legitimately stamps with `updatedAt`). Paired with a hard "you may not edit
+tests" bind and no escalation path, cheating was the only route to green. This is
+shape 9's lesson again: **every hard bind needs an escalation path**, and a spec that
+forbids editing tests must add "if a test looks unsatisfiable, STOP and name it in
+notes.md — a reported conflict is an acceptable outcome."
+
+**Routing takeaway.** Do not hand this model a task whose success condition is a
+frozen suite you have not yourself proven satisfiable. Either prove the suite passable
+first, or give it an explicit escalation path — preferably both. The orchestrator
+hand-authored the module after the second rejection (operator ruling), and the same
+suite went green with no test-awareness at all, which confirms the suite — once
+repaired — was the variable, not the model's capability.