Note ipywidgets cache monkeypatch work and in-app measurement caveat

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: maartenbreddels

notes.md

@@ -368,6 +368,44 @@ Implemented and E2E-validated (playwright headless):
   version/cache-busting story before defaulting on. VueComponent (component_url) is the
   symmetric next step (httpVueLoader URL mode already exists in the JS).
 
+## ipywidgets caches as a monkeypatch in exploration repo (2026-06-15)
+
+Goal: ship the ipywidgets get_state caches without waiting on upstream. Confirmed our
+perf-widget-construction branch is a SUPERSET of Maarten's stalled PR #3602 (it has #3602's
+keys-cache + trait_to_json-cache, PLUS a static-default fast path, and deliberately OMITS the
+breaking comm-trait removal that blocked #3602 — so a monkeypatch is safe).
+
+Built in a git worktree (exploration-wt-ipyw, branch perf/ipywidgets-cache-monkeypatch, commit
+9bf4b44b8, UNSIGNED — gpg couldn't prompt; local only, no PR):
+- leasing/src/shared/utils/ipywidgets_cache.py — default-on monkeypatch of Widget.get_state +
+  per-class caches, with a per-class verify-on-first-use guard (falls back to stock per class on
+  any byte mismatch).
+- leasing/tests/shared/test_ipywidgets_cache.py — 35-case offline byte-identical test (passes vs
+  stock 8.1.8). Invariant: patched output == pre-install output, so meaningful in any venv.
+- wired install() into solara_app.py before install_perf_hooks (before any widget exists).
+
+TWO REAL BUGS found only because reconfirmation was insisted on (first 'PASSED' was bogus —
+stock-vs-stock for disabled classes):
+1. Reentrancy: _verify_class constructs a probe -> __init__ -> open() -> get_state() re-entered
+   verification for the same class -> class wrongly disabled (+ the RecursionError spew). Fixed
+   with a _verifying sentinel.
+2. Two-instance compare: probes embedding layout/style sub-widgets have per-instance model ids
+   that never match -> all core widgets disabled. Fixed by verifying on a single instance.
+
+Delta reconfirmed (stock vs patched, identical-deps venv): widget create 1.33-1.45x; 200 v.Btn
+construction 24.5->18.7ms (1.31x); byte-identical across 14 widget types + add_traits; guard
+verifies all (none disabled).
+
+IN-APP caveat (couldn't run real leasing app — needs secrets/DB): measured proxies only.
+- synthetic 1024-button solara render cycle: only ~3% faster (952->930ms min) because that bench
+  is reacton-reconciler/comm-dominated; widget construction is a small fraction of it.
+- the leasing home page is get_state-DOMINATED (other agent profiled get_state at 41% of the
+  cycle, cut 0.383->0.171s, render cycle -28% with the equivalent editable branch). The monkeypatch
+  is byte-identical to that branch with the same microbench speedup, so the -28% should transfer.
+- DEFINITIVE in-app number needs the leasing agent: force-reinstall stock ipywidgets==8.1.8 in the
+  leasing venv (it still has the editable branch, now redundant — and the monkeypatch would
+  otherwise wrap the branch), run from the worktree, confirm with PROFILE=1.
+
 ## Stash (possible bugs / questionable behavior found along the way)
 
 - pandas 3.0 re-render on equal DataFrame (see above) — env pin papering over it for now.