test: cap _SiteStats.record shape and caller sets (PRPUNDIT-13)

[jiagaoxiang]

Summary

• Calls _SiteStats.record past _MAX_CONTAINER_WIDTH / _MAX_CALLERS_PER_SITE and asserts count/wall/nbytes and first/last keep moving while the sets stop growing.

Closes test gap PRPUNDIT-13.

Test plan

• PYTHONPATH=src pytest src/hyperloom/inference_optimizer/tests/test_framework_rewrite_evidence.py::test_site_stats_record_caps_shape_sigs_and_callers_while_tallies_grow

[zoroyihan7]

Thanks — the caps are solidly pinned: pushing past both bounds, asserting the counters keep accumulating, and explicitly asserting the overflow entries are absent from the sets is exactly the right shape, and removing either len(...) < CAP condition turns this test red.

One gap: first_s is only half-pinned. Mutating the sentinel init at hl_host_probe.py:223:

self.first_s = 0.0      # was: -1.0

kills the latch outright — if self.first_s < 0 never fires again, so first_s is permanently 0.0 and never records the real first timestamp — and all 73 tests stay green.

The reason is that the loop's first call passes at_s=0.0, which is exactly the degenerate value, so assert stats.first_s == 0.0 can't distinguish "latched the first timestamp" from "never latched anything". (Deleting the if so first_s is overwritten every call is caught — that's the other half.)

One-token fix: drive the loop with at_s=float(i) + 1.0 and assert first_s == 1.0. That kills both the sentinel mutation and the overwrite mutation.

Two nits while you're in here:

• The falsy-input branch of the guards (if shape_sig and ... / if caller and ..., lines 242/244) isn't pinned — replacing them with bare if len(...) < CAP: leaves the file green. Passing shape_sig="" / caller="" and asserting the empty string is absent would be two more lines. Outside the contract this PR claims, so genuinely optional.
• Re-asserting first_s after the post-cap record would make the latch/overwrite contrast explicit next to the last_s == 99.0 assertion.

[jiagaoxiang]

at_s=float(i)+1.0 with first_s == 1.0 pins the sentinel-init mutation. Re-asserted first_s == 1.0 after the post-cap last_s == 99.0 record, and added a case that records empty shape_sig/caller while the sets still have room.

[zoroyihan7]

Re-reviewed. The at_s offset closes the sentinel hole.

Mutations red now:

• self.first_s = -1.0 → 0.0 in _SiteStats.__init__ (L223), which kills the latch outright — red (2 tests)
• if self.first_s < 0: → always true, i.e. overwrite on every call — red
• if shape_sig and guard dropped (L242) — red
• if caller and guard dropped (L244) — red

Both halves of the latch contract are pinned now. 74 passed. LGTM.

FYI, not for this PR: the identical sentinel at L290 belongs to a different class (depth / arg_samples / strict_sigs) and stays green under the same mutation. Might be worth its own ticket.

[jiagaoxiang]

Leaving the _CallStats sentinel (hl_host_probe.py:290) out of this PR as discussed. Filed #1356.