perf(handler): reduce render lookup and scope allocations

[Mohamed Mansour (mohamedmansour)]

Verdict

GO as one combined performance change. This PR now contains both the render-fragment index and borrowed render scopes. It consolidates and supersedes #474.

Summary

• precompute a compact render-fragment index once in Protocol::new, replacing repeated fragment-target hashes, component prop-name conversion, and unnecessary sibling-route scans with numeric slots and flat metadata arenas
• borrow state-backed loop items and component attributes through ordinary render scopes instead of cloning JSON subtrees
• preserve lexical precedence across nested loops, component locals, and global fallback
• materialize borrowed component attributes only where progressive streaming must retain them across host calls
• keep all new APIs crate-private; no protobuf, public API, or unsafe changes

Combined performance case

B1 alone moved work from render time to protocol construction. It improved representative fragment-heavy renders by roughly 2–5%, but added a one-time Protocol::new cost of +24 µs for 128 fragments and +128 µs for 512 fragments, plus +3.6–5.1% retained protocol memory. It also had one measured all-miss condition case at +4.15%.

B2 makes the combined stack clearly favorable. Sequential custom-allocator measurements below compare B1 with this combined B1+B2 branch on the contact-book /contacts route:

Contacts	FAST	Allocations	Allocated bytes	Median render	Throughput
10	no	814 → 220 (-73.0%)	36,143 → 9,829 (-72.8%)	58.855 → 36.432 µs (-38.1%)	+61.5%
10	yes	982 → 388 (-60.5%)	86,420 → 60,106 (-30.4%)	84.353 → 62.702 µs (-25.7%)	+34.5%
100	no	6,934 → 1,210 (-82.5%)	277,972 → 33,049 (-88.1%)	466.528 → 258.856 µs (-44.5%)	+80.2%
100	yes	7,104 → 1,380 (-80.6%)	426,553 → 181,630 (-57.4%)	560.254 → 355.026 µs (-36.6%)	+57.8%
1,000	no	68,134 → 11,110 (-83.7%)	2,713,293 → 267,949 (-90.1%)	5.038 → 2.491 ms (-50.6%)	+102.3%
1,000	yes	68,308 → 11,284 (-83.5%)	4,827,954 → 2,382,610 (-50.6%)	6.434 → 3.789 ms (-41.1%)	+69.8%

Allocation deltas are deterministic: exactly 594, 5,724, and 57,024 allocation calls are removed at the 10/100/1,000 scales in both plugin modes. Timing is corroboration; the allocation counts are the primary evidence.

B2 does not hide B1's construction tradeoff. All 16 construction fixtures retain exactly the same bytes under B1 and combined B1+B2 (contact-book remains 102,619 bytes; the 200-route synthetic remains 574,831 bytes). The combined render savings amortize B1's largest measured construction penalty in one render at 100/1,000 contacts and in roughly six renders at 10 contacts. Completed renders retain zero additional bytes.

For the isolated B1 layer, minimum-time measurements against the original base recorded loop_scaling/500 -5.54%, plugin_fast/without_plugin -4.89%, and loop_scaling/100 -4.58%. The combined measurements above then show the larger B2 gain on top of B1.

Correctness

• 90 deterministic RENDER/STREAM harness lines are byte-identical between B1 and combined B1+B2
• excluded only the documented nondeterministic component-assets path=/ plugin=true line, which flips between the same two hashes on the original base
• retained protocol bytes are unchanged between B1 and combined B1+B2 for all 16 fixtures
• resolver tests cover borrowed global collections, borrowed component collections, nested loops, owned-local precedence, lexical shadowing, and component isolation
• progressive streaming keeps owned continuation state and avoids a duplicate resolver lookup

Implementation notes

RenderFragmentIndex reuses interned fragment IDs, stores metadata and canonical component prop names in flat arenas, tracks route presence in a bitset, and resolves fragment lists lazily through eight inline memo slots plus an on-demand spill. Render and continuation slots use the same sorted fragment-ID numbering.

Ordinary rendering uses BorrowedScope, LoopBinding, and VisibleLoopScope to retain references into immutable request state. Four inline borrowed-scope slots cover the common component-attribute case; overflow storage and request-local pools remain bounded. Owned fallback paths preserve synthetic values and continuation lifetimes.

Validation

• cargo test -p microsoft-webui-handler
• cargo xtask check — all phases passed, including clippy, deny, workspace tests, native/WASM builds, examples, benchmark validation, and docs (using same-drive TEMP/TMP required by the Windows projection adapter)