proto(#262): Phase-3 measurement — slot model = value path for all predicates

Measured the first Phase-3 question (is the assignment-driven slot model
viable as the single model?) by extending slot coverage to name-based
(global) bindings and running the suite flag-on.

Result: 45 failures, ALL of the form "got 0 expected 1" — full-window
predicates returning false because the slot window is exactly 1 short. Not
a sampling-regime divergence: the slot model matches the value path for
every predicate. The sole gap is a 1-count FIRST-ASSIGNMENT lag on name
bindings — the binding is created by the upcoming SET_NAME and does not
exist at OBSERVE_ASSIGN time, so the slot starts at the 2nd assignment.
(Function-locals are unaffected: their slots are pre-allocated by the
compiler, so they observe from assignment 1 — Phase 2 already green.)

So the assignment-driven slot model is sound; Phase 3 needs first-assignment
coverage for name bindings (observe-at-SET, or slot-promote interrogated
module vars), then report on the single model, then delete the value path.
Exploratory name-based shadow reverted to keep Phase 2 flag-on green; the
finding is recorded in the OBSERVE_ASSIGN comment.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: InauguralPhysicist

src/vm.c

@@ -3664,9 +3664,14 @@ static Value *vm_run(EigsChunk *chunk, Env *env) {
                 }
                 v->dirty = 1;
                 g_last_observer = v;
-                /* #262 Phase-1: a name-based observe is not slot-shadowed —
-                 * invalidate the slot override so the next predicate uses the
-                 * authoritative value path. */
+                /* Name-based observe is not slot-shadowed (yet) — invalidate
+                 * the slot override so the next predicate uses the value path.
+                 * Phase-3 measurement showed shadowing here works for every
+                 * predicate EXCEPT a 1-count first-assignment lag (the binding
+                 * is created by the upcoming SET_NAME and doesn't exist yet);
+                 * fixing that needs first-assignment coverage (post-set observe
+                 * or slot-promoting interrogated module vars), tracked for
+                 * Phase 3. */
                 if (obs_shadow_on()) g_last_obs_slot_idx = -1;
             }
         }