perf: avoid cloning nested route trees

[Mohamed Mansour (mohamedmansour)]

Matched route descent deep-cloned the entire remaining WebUiFragmentRoute subtree. Because each route recursively owns strings, child vectors, cache tags, and invalidation data, the cost grows with route depth and sibling count on every request.

This change keeps protocol-owned route levels borrowed during ordinary rendering and materializes them only when a streaming continuation must outlive that borrow.

Approach

• Represent the active route level as Cow<'protocol, [WebUiFragmentRoute]>.
• Borrow child levels directly while descending protocol-owned routes.
• Keep parked SessionCore and ContinuationVm route levels owned across host calls.
• Move children out of an already-owned route with mem::take instead of recursively cloning them.
• Preserve the existing second-<outlet /> behavior exactly; its latent bug remains isolated in [Bug]: Second <outlet /> at the same route level silently renders nothing #515.

The scope maps and name sets deliberately remain String-keyed, and the scope-map pool remains live across suspension steps. An earlier broad Cow<str> prototype improved one-step rendering but a real three-boundary session exposed a deterministic regression: 243 to 249 allocations and 25,081 to 29,539 allocated bytes per run. That design was rejected rather than accepting a streaming trade-off.

Exact allocation results

Baseline is the current stacked base, #511 at e908d327. Counts come from the benchmark counting allocator and were deterministic across repeated runs.

streaming_resource_bench, scale 1000:

Path	Before allocs / bytes	After allocs / bytes	Output
string	141 / 31,237	127 / 29,488	24,152 B
streaming one-step	155 / 44,637	141 / 42,888	24,152 B
streaming one-step POOLED	148 / 8,893	134 / 7,144	24,152 B

The same 14-allocation and 1,749-byte reduction holds at scales 10 and 100. Output remains byte-identical at all scales: 24,114 / 24,134 / 24,152 B.

Nested routes:

Request path	Before allocs / bytes	After allocs / bytes	Output
/	58 / 8,418	49 / 7,662	3,366 B
/sections/a	86 / 9,944	77 / 9,188	3,902 B
/sections/a/topics/b	107 / 11,388	98 / 10,632	4,455 B
/sections/a/topics/b/lessons/c	138 / 18,121	129 / 17,365	4,740 B

Each route output length and checksum is identical before and after.

Suspension regression checks

The existing benchmark's streaming rows do not encounter a runtime boundary, so they cannot validate state parked across host calls. #517 tracks adding a durable suspending case. This PR was additionally measured with two real StreamingSession probes:

Session path	Before allocs / bytes	After allocs / bytes	Output
Three boundaries with populated scopes	243 / 25,081	243 / 25,081	4,940 B
VM-owned nested routes delegated to ordinary outlets	233 / 24,687	233 / 24,687	2,144 B

Both paths exactly match the base for allocations, allocated bytes, and output. The nested-route probe specifically covers the borrowed-to-owned transition and caught an intermediate recursive clone before the final move-based implementation restored exact parity.

Wall time

No speedup is claimed. Separately built release binaries were run in alternating order, n=9 per build and path, with warm-ups. Every distribution overlaps, so the defensible result is no measurable wall-time regression.

Path, wall us/run	Before median [range]	After median [range]
resource string/1000	13.02 [12.87-13.76]	12.85 [12.52-13.15]
resource POOLED/1000	13.64 [13.56-13.87]	13.49 [13.08-13.80]
deepest ordinary route	7.300 [7.237-7.391]	7.179 [7.096-7.350]
VM-owned nested streaming route	11.013 [10.906-11.092]	11.056 [10.990-11.623]

RSS is intentionally omitted from the comparison because #516 demonstrates that the current column is multimodal and dominated by fixture construction. Exact allocator counts are unaffected by that issue.

Validation

• cargo test -p microsoft-webui-handler: 457 unit and 36 integration tests passed.
• cargo xtask check: full workspace gate passed.
• Added tests for borrowed pointer identity, move-based owned descent, owned session parking, scope-pool retention across suspension, and the ContinuationVm to ordinary-render nested outlet path.
• Independent performance review found no remaining correctness, lifetime, or allocation regression.

No public API or behavioral contract changed, so no DESIGN.md or developer-doc update is required.

---

Stacked on #511 and targeting mohamedmansour-reduce-handler-render-allocations. Review #511 first.

[Copilot]

🟢 Approval recommended

The ownership transitions are coherent, preserve behavior, and have focused regression coverage.

Pull request overview

Avoids deep-cloning nested route subtrees during handler rendering while preserving streaming ownership requirements and existing behavior.

Changes:

• Represents active route levels with Cow.
• Borrows protocol routes and moves already-owned descendants.
• Adds ownership, suspension, and nested-route regression tests.

File summaries

File	Description
crates/webui-handler/src/lib.rs	Implements borrowed and move-based route descent.
crates/webui-handler/src/streaming/vm.rs	Preserves owned routes in continuation frames.
crates/webui-handler/src/streaming/session.rs	Parks route levels as owned session state and tests retention.
crates/webui-handler/src/streaming/inventory.rs	Updates test context initialization for Cow.
crates/webui-handler/tests/streaming.rs	Tests nested owned-route descent through streaming rendering.

Review details

• Files reviewed: 5/5 changed files
• Comments generated: 0
• Review effort level: Balanced

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.