perf(mcp): reduce global dispatch catalog construction churn

[ScriptedAlchemy]

Separate from #810: the typed RMCP connection benchmark now measures steady-state dispatch only after explicitly warming the process-global dispatch catalog.

Evidence

A fresh mcp.server.construct Hotpath measurement on the eager-construction experiment recorded 505,598,728 allocated bytes across 3,762,998 allocations (one constructor call). The catalog is a process-global LazyLock; eagerly resolving it in McpServer::new_with_context changes when this work is paid, not its size.

A hermetic project-scoped production-MCP construction-only fixture opened the server, did not bind a listener or establish an RMCP connection, retained it for two seconds for process sampling, then shut it down. It compared 9adc7c4 with the unpublished eager experiment d185c420:

metric	9adc baseline	eager experiment	delta
profiler wall	3.512 s	4.453 s	+0.941 s
process HWM	192,748 KiB	222,468 KiB	+29,720 KiB
retained RSS	182,939,648 B	221,360,128 B	+38,420,480 B
retained-RSS delta	72,757,248 B	112,431,104 B	+39,673,856 B

The eager experiment is deliberately not being integrated. It demonstrates the lifecycle cost that a real reduction must avoid, not a fix.

Acceptance

• Localize the dominant allocations inside catalog construction and reduce the actual construction work; do not merely prewarm or move it between first dispatch and project open.
• Keep a no-RMCP project-open journey free of catalog construction. Measure its wall time, retained RSS, and HWM with the same hermetic process fixture, and report first real RMCP use separately.
• Preserve the typed dispatch route, cancellation, selected-project leases, delivery settlement, and response identity; verify the real broker/RMCP wire journey rather than a provider lookalike.
• Keep the steady-state benchmark explicit about its prewarmed catalog scope so future dispatch-byte numbers cannot be interpreted as cold process-start costs.

No live TraceDecay profile or daemon was used for these measurements.

[ScriptedAlchemy]

Implemented and pushed on agent/issue-861-catalog-construction at 5795babf2 (three commits: durable no-RMCP fixture, MCP catalog borrowing, remaining HTTP/SDK borrowing).

Root cause and cold first-use result

Matched real broker/RMCP Hotpath runs localized the churn to repeated deep clones of schema-rich process-static registries while assembling the one global dispatch catalog:

• Work registry clone: 66 calls, 443.6 MB exclusive allocation (71.56% of the baseline report)
• Workflow registry clone: 2 calls, 5.5 MB
• Work lookup nested allocation: 64 calls, 13.4 MB

The accessors now return borrowed process-static authorities. MCP per-tool lookups and the intermediate dispatch catalog retain borrowed executable bindings. The same avoidable clones were removed from Work/Workflow HTTP request dispatch and SDK aggregation with Cow<'static, ExecutableBindingRegistryV1>; constructors that genuinely return owned registries remain owned.

Cold first-RMCP-use comparison (same pre-#810 benchmark scope, so this includes the one lazy catalog build):

metric	before	fixed	delta
Hotpath exclusive function allocation	619.8 MB	137.8 MB	-482.0 MB (-77.8%)
mcp.server.rmcp.dispatch allocated bytes / 143 dispatches	10,346,241 B	230,097 B	-97.78%
Hotpath elapsed	2.41 s	1.30 s	-1.11 s

The 443.6 MB Work and 5.5 MB Workflow clone spans disappear completely from the fixed report. Reports: /tmp/tracedecay-861-before.json and /tmp/tracedecay-861-after.json; fixed ticket cc-4667.

No-RMCP constructor lifecycle

Commit 2a54b63b6 adds a durable hermetic fixture using the real project-scoped runtime, graph initialization, and production McpServer constructor. It deliberately binds no listener and sends no RMCP request, retains the server for a fixed two-second sampling interval, then shuts it down. The process-global dispatch catalog remains lazy.

Matched fixture-only 2a54b63b6 versus final 5795babf2:

metric	fixture-only baseline	fixed	delta
OS collector wall (includes fixed 2 s hold)	3.373 s	3.410 s	+0.037 s
retained RSS	169,832,448 B	161,796,096 B	-8,036,352 B
GNU-time wall (separate matched HWM run; includes hold)	3.69 s	3.58 s	-0.11 s
maximum RSS/HWM	154,160 KiB	144,504 KiB	-9,656 KiB

Reports: /tmp/tracedecay-861-construction-baseline/report.json, /tmp/tracedecay-861-construction-fixed-final/report.json, and matching *-time.txt. Fixture builds cc-4689, cc-4726.

Separate #810 steady-state guardrail

This is not used as the construction result. With #810's explicit catalog prewarm applied, cc-4698 reproduced its central result exactly: 24,439 B / 143 = 170.902 B per dispatch. After the final HTTP/SDK borrowing follow-up, cc-4735 measured 18,471 B / 143 = 129.168 B per dispatch. Thus the cold-construction reduction does not hide a steady-state regression; #810 retains authority for the prewarmed benchmark scope.

Behavioral verification

• Borrowed-authority regressions: cc-4752, cc-4755, cc-4792, cc-4793
• Canonical dispatch catalog: cc-4786
• Real broker/RMCP wire matrix: cc-4787
• Connection cancellation identity, selected-project retirement, response/RPC identity: cc-4789, cc-4790, cc-4791
• Strict Clippy: application cc-4797; composition plus RMCP fixture cc-4801
• Commitlint, rustfmt, and git diff --check: clean
• Independent rereview at exact 5795babf2: clean; no Critical or Warning findings. It confirmed lazy/error/lifetime semantics and unchanged dispatch, cancellation, lease, delivery, and identity behavior.

No live TraceDecay profile, daemon, CLI, or MCP service was used.

[ScriptedAlchemy]

Merged down onto PR #707 at efce469 (merge commits e042f3b codex/integrate-sept4, 4bef15a codex/issue-863-ci-isolation-final, d0da2f4 codex/fix-post-trim-pressure-state, efce469 codex/port-753-activation-net). The combined tree passes cargo check --workspace --all-targets --locked (hauler cc-5021, 0 errors, 0 warnings); the workflow conflict between the two Windows-archive copies was resolved toward the 863 lane. Hosted CI on this tip is the next acceptance signal.

[ScriptedAlchemy]

Closed by central PR #707 at efce469: agent/issue-861-catalog-construction (2a54b63 no-RMCP construction fixture, d228015 borrow global dispatch catalog authorities, 5795bab borrow static HTTP registry authorities) is merged down via codex/integrate-sept4. Hosted CI on the merged tip is run 33915402108; reopen if the cold first-use measurement regresses there.