[AGiorgetti]

Shallow ThrowOnCycle mappings eagerly allocate a HashSet even when only one source is active. Store that first source in one object slot, then promote to reference-identity spill storage for a second simultaneous source and retain it across siblings. Empty/null and reserved sentinel/type-tag sources remain distinct; cycle paths, exception metadata, Enter/finally/Exit cleanup and static generated mapping are unchanged.

Adds focused state, collision, identity, failure cleanup, reentrancy and allocation tests, including both simultaneous null/self orderings. Updates existing storage assertions and records measured performance limits. No public API, consumer-skill contract, benchmark harness, threshold or runtime policy change.

Validation

• Release solution restore/build: zero warnings/errors. Independent seven-file review: no blocking findings.
• Final production core tests: 1,056 generator, five runtime and one integration tests pass, zero failures/skips. All 273 selected local package tests also pass, zero failures/skips, including API compatibility, framework/language consumers, samples and trimming. Native AOT and benchmark smoke are the two explicitly excluded local cases. Exact-head CI 38405298319 passed: all 1,337 solution tests on each Windows/Linux host and 1,056 generator tests in each of six Roslyn 4.8/4.14/5.9 jobs, zero failures/skips. The full package suite includes API compatibility, framework/language consumers, trimming, required Native AOT and benchmark smoke. PR merge tree equals the reviewed head tree. The three release-only jobs were skipped by design; no final-version release artifacts or publication were produced.
• Local benchmark smoke is deferred because its default runner can switch power plans. Native AOT is deferred to the required CI toolchain; MSVC discovery found no local installation. No missing local tools were installed and no persistent machine settings were changed.

Performance and release boundary

On SPECTRE_GRD, tracked depth-1 allocations fell from 232 to 56 B/op; depths 2/16/128 remain 264/1,744/15,528 B/op. The six-pass three-design timing experiment had control drift up to 94.14%, reversed compact/inline-first shallow ordering, and deeper slowdown signals. All four compact/develop gate comparisons blocked. This establishes an allocation improvement, not an overall throughput win or release performance pass; the earlier 3.0.0/develop shallow timing concern remains unresolved. No measurements from other PCs were pooled.

Full raw evidence: compact experiment archive, SHA256 1D78057FA23247A5A3637132A9B9A5FFBE0A4E445C4C7F4A91588EDA6B995F59. Stable same-binary controls and a final-version nonpublishing release rehearsal remain release gates. Draft for human review; no merge, tag or publication is requested.