alkhttp/tasks/adapters/review-002-oai11-ref-memoization.md

id, name, status, depends_on, scope, risk, impact, level, tags

id	name	status	depends_on	scope	risk	impact	level	tags
review-002-oai11-ref-memoization	Memoize $ref expansion to bound resolver work (OAI-11)	completed		moderate	medium	component	implementation	adapters review-002 from-openapi openapi-spec

Description

Review 002 OAI-11 [major, borderline critical]. The OAI-01 fix (cycle guard + depth budget, openapi_spec.rs:382-437) bounds stack, not work: every $ref hop re-expands its target from a fresh clone with no memoization. A chain of shared but acyclic refs —

S_i = { a: { $ref: "S_{i+1}" }, b: { $ref: "S_{i+1}" } }

(the doubly-linked-list prev/next shape — valid, common) expands as a full binary tree. Empirically (verbatim harness): 20 levels → 7.3M node visits / 3.6 s; growth ≈ 2^levels → 30 levels ≈ ~1 hour, 40 ≈ days. Neither guard fires (err=false); the import just wedges — no stack overflow, no error. The same "valid document kills the process" class as OAI-01, one mitigation shy.

Acceptance Criteria

• resolve_refs_bounded memoizes resolved, cycle-free $ref targets per document (cache keyed on the ref string; clone the resolved Value on hit instead of re-resolving) — or an equivalent total-node-count budget with a clean SchemaParse-family error (implementer's choice; memoization strongly preferred — it preserves full acyclic support)
• The cycle guard and both depth checks remain (memoization must not alter cycle detection semantics — a cycle still errors)
• Acceptance gate: a 30+ level shared-chain spec imports in bounded time (add the test with a wall-clock or node-count assert; a 30-level chain must import in < 1 s)
• Tests: diamond (non-cyclic shared refs) unchanged-correct; cycle-through-shared-node still errors; memoized result equals the old expansion for a shared-ref schema (golden compare on one case)
• The four missing cycle-shape tests from OAI-16 (array-items, allOf, additionalProperties, $ref-sibling) land here too — same function, same risk envelope
• cargo test, cargo clippy --all-targets -- -D warnings, cargo fmt --check pass

References

• docs/reviews/002-post-remediation-review.md (Part E', OAI-11 + OAI-16)
• src/adapters/openapi_spec.rs:382-437 (the resolver), :713-805 (existing cycle tests), from_openapi.rs:736 (the requestBody-chain test)
• tasks/adapters/review-001-ref-cycle-guard.md (the OAI-01 fix this completes)
• docs/architecture/decisions/066-from-jsonschema-as-http-adapter.md

Notes

Clone cost per hit is fine (schemas are small; the win is avoiding exponential re-resolution, not allocation). If memoizing, note the resolving-set interaction: a memo entry must only be written for completed (cycle-free) expansions — a partial expansion under an active cycle attempt must never be cached (write-on-success at the top of the recursion return path, or post-order). This is the one subtle part; the existing cycle tests + the four new shape tests are the safety net.

Summary

Memoized $ref expansion (write-on-success cache) + node-budget accounting on memo hits + separate MAX_REF_HOP_DEPTH (64) from structural depth (128). Exponential diamond chains now fail with a clean budget error (<1s) instead of OOM. Note: the original WIP's unbounded memo-hit clone OOM-killed the dev server twice during implementation; verified under a 2G systemd scope.