wcc_ww/cgen: c3 in-file peel sites — exprprimresolved base walk chases (B1); elemsizeofc alias-indexable leg (B3); B2 probe-cleared

F2a batch-2 c3, probe-decided per rob's spec.

B1 (exprprimresolved N_DOT, was :2073/:2075/:2078): the hand-rolled
NAMED->PTR->NAMED->STRUCT walk ran out at a 3-level alias base or a
ptr-to-2-level base — the cast-source width stayed unknowable on
wwstage only, so the #33 identity clamp was emitted where cstage
skipped it (runtime-correct both, byte-id NO, gate-blind). cs twin
castsrcprim is F1-chased at both hops (type_chase_named) -> ww-only
align-up, no fuse. Probes: b1b_3lvl + b1b_ptr2lvl graduate 0/0
byte-id; 1/2-level controls held throughout (the 2-peel covered
2 levels by accident); the FIELD-u chase is asm-neutral (cs keeps a
single peel there, sound via type_isint NAMED recursion + NAMED size
carrying the underlying's — probe b1c_fld2lvl byte-id). 944 rows
castprim_3lvl_base / castprim_ptr2lvl_base / castprim_2lvl_base_ctl.

B2 (nullableptrtag variant scan :2719): PROBE-CLEARED, NO FIX. The
2-level ptr-alias nullable variant (b|void), the order-swapped
(void|b) that defeats the return-0 fallback, AND 3-level all run 0/0
byte-id both stages — the constructible variant params never carry a
2+-level NAMED at this scan. cs twin nullable_ptr_tag (cgen.c:747)
has the identical single peel: parity, both correct on every
constructible shape. NOT nominal-sensitive (shape-keyed first-PTR
scan, no name choice exists) — the spec's #209/#211 hold condition
does not trip. The single-peel if stays, classified survivor.

B3 (elemsizeofc): an alias-NAMED INDEXABLE (`type grid = [3]cell`)
arrives as a bare N_TNAME — elemsizeof's name arm knows only
str/prims and answered the 1-sentinel through the `direct != 8`
short-circuit (#60 esz-1 family, outer-array leg). New N_TNAME leg
answers from the chased stamped tinfo via idxeffti. #83 caller
probes (rule: no unprobed caller row):
- append leg: UNREACHABLE on ww — `let xs: nums = [];` rejected by
  the wwstage CHECKER ("let: not assignable") even at 1 level; cs
  accepts+runs. check.ww is batch-4 scope -> documented on #83.
- alloc-empty leg: BOTH stages loud-reject the alias-typed binding
  hint ("cannot infer slice element type", check.ww:2547 + cs twin)
  — symmetric loud, alias-blind hint documented on #83.
- arr[i].field= float leg: plain [3]cell runs 0/0 byte-id; the
  alias-named outer array LOUDS on ww at its own site
  (cgenexpr.ww:11033 assign-resolver float-field bound), cs runs —
  fails upstream of elemsizeofc, documented on #83.
- subslice-of-alias-array call-arg probe: already 0/0 pre-change
  (the #60/#257 stamped-tinfo keying answers first) — the internal
  chase is presently LATENT defense-in-depth; ZERO callers enroll.
Remaining elemsizeofc internal loops (eti/esub) fold to tichase.

Byte-id: five-mains NEUTRAL vs the c2 build on identical inputs;
cs==ww on the regenerated five; 944 138/138; test-unit 288 green.
combined.ww regens ride along.

test/wcc/944_alias_accept_run.c

@@ -753,6 +753,50 @@ static const struct row rows[] = {
 	  "	};\n"
 	  "	return 0;\n"
 	  "};\n", 0, K_RUN, NULL },
+	/* ---- F2a batch-2 c3/B1: exprprimresolved's N_DOT base walk was
+	 * a hand-rolled 2-peel — a 3-level alias base (or ptr-to-2-level)
+	 * left the cast-source width unknowable on wwstage only, so the
+	 * #33 identity clamp stayed emitted where cstage (castsrcprim,
+	 * F1-chased) skipped it: runtime-correct both, byte-id NO.
+	 * Graduated by the tichase walk; 1/2-level controls held all
+	 * along (the 2-peel covered them by accident). */
+	{ "castprim_3lvl_base",
+	  "package main;\n"
+	  "type s1 = struct { f: u32 };\n"
+	  "type s2 = s1;\n"
+	  "type s3 = s2;\n"
+	  "export fn main() i32 = {\n"
+	  "	let x: s3;\n"
+	  "	x.f = 70000;\n"
+	  "	let y: u64 = (x.f: u32): u64;\n"
+	  "	if (y != 70000) { return 1; };\n"
+	  "	return 0;\n"
+	  "};\n", 0, K_RUN, NULL },
+	{ "castprim_ptr2lvl_base",
+	  "package main;\n"
+	  "type s1 = struct { f: u32 };\n"
+	  "type s2 = s1;\n"
+	  "export fn main() i32 = {\n"
+	  "	let x: s2;\n"
+	  "	x.f = 70000;\n"
+	  "	let p: *s2 = &x;\n"
+	  "	let y: u64 = (p.f: u32): u64;\n"
+	  "	if (y != 70000) { return 1; };\n"
+	  "	return 0;\n"
+	  "};\n", 0, K_RUN, NULL },
+	/* 2-level control (was already byte-id pre-c3; pins the accident
+	 * out of existence). */
+	{ "castprim_2lvl_base_ctl",
+	  "package main;\n"
+	  "type s1 = struct { f: u32 };\n"
+	  "type s2 = s1;\n"
+	  "export fn main() i32 = {\n"
+	  "	let x: s2;\n"
+	  "	x.f = 70000;\n"
+	  "	let y: u64 = (x.f: u32): u64;\n"
+	  "	if (y != 70000) { return 1; };\n"
+	  "	return 0;\n"
+	  "};\n", 0, K_RUN, NULL },
 	/* ---- task #71 (rob probe-ruled F1 enrollment): nested fields
 	 * whose TYPES are aliases. The chained-dot store/read/addr-of
 	 * walks single-peeled each hop, aborted on the alias, and fell