repsel: --opt-report panics on real dependency JS — RULE2 carries two tiers and render.rs asserts one

[proggeramlug]

--opt-report panics on real dependency JS in any debug-assertions build,
because RULE2 ("rule 2 (containment)") carries two tiers in
collectors/ptr_shape_report.rs — 10 records with Tier::Fixable and 3 with
Tier::CompilerLimitation — and opt_report/render.rs:418 asserts that a rule
carries exactly one.

thread 'perry-main' panicked at crates/perry-codegen/src/opt_report/render.rs:418:9:
assertion `left == right` failed: rule rule 2 (containment) carries two tiers; the bucket would report only one
  left: Some(CompilerLimitation)
 right: Some(Fixable)
Error: perry compiler panicked: …

The assertion is #7176's own S20 guard (one_rule_carrying_two_tiers_is_a_hard_error),
added because a debug_assert nothing exercised was a green hole. It is
correct; the table it guards is not.

Reproduction

cargo build -p perry                       # debug — assertions ON
mkdir /tmp/r && cp <module> /tmp/r/mod.js && cd /tmp/r
PERRY_NO_AUTO_OPTIMIZE=1 <perry> compile mod.js --opt-report=json --no-link

Module: @typescript-eslint/eslint-plugin/dist/rules/naming-convention-utils/schema.js
(from real-apps/scriptc/node_modules). It emits both a Fixable rule-2
denial and a CompilerLimitation one, so the renderer's per-rule tier fold
sees two.

Why nobody has hit it

• --profile perry-dev inherits release, so debug_asserts are off —
  every local repsel measurement in this campaign has been perry-dev and never
  saw it.
• The repsel-census CI job DOES build cargo build -p perry (debug), but its
  corpus is benchmarks/repsel_census + benchmarks/suite, and no fixture
  there lands both tiers under rule 2. Real dependency JS does, routinely.

Scope

Found while measuring #7170 R1 (PR #7233), and exonerated by in-place A/B:
with all five of that PR's production files reverted to origin/main in the
same target dir and the same package set, the panic is identical. It is
pre-existing at 46e5e377b, and it is not introduced by the three GC commits
between 65a64f428 and there — reverting to origin/main sources is what
reproduces it.

Fix shape (two options, not equivalent)

1. Split the rule string. If a reassigned binding and a delete really are
   different tiers, they should be different rule strings, and
   ALLOC_BUCKET_FLOORS / by_rule would then tell them apart — which is the
   whole point of perf(repsel): make the Ptr<Shape> alloc-site buckets mean what schedulers read them as (#7170 R0) #7176 §5's composite key.
2. Pick one tier for rule 2. Cheaper, and loses the distinction the records
   were written to carry.

Either way the fix needs a test that puts both tiers in one report, because
that is the case no existing fixture reaches — the same gap that let this ship.