Add opt-in observations for distributed action costs

[alexy]

Purpose

Measure the existing distributed action path for Sem's C2 study. The observer is opt-in (telemetry.c2_observer=false by default) and adds no validation or dataset actions.

Changes

• Typed identities and outcomes for 43 named phase observations across planning, serialization, scheduling, stream ownership, output and cleanup.
• Inclusive intervals, process-local sequence and counter summaries, and explicit failure, cancellation and abandonment outcomes.
• Stream delegation and response-buffer ownership observations with sibling tests for disabled behavior, errors, EOF and drop.
• Enabled-only RPC context for release and worker registration, preserving disabled lazy behavior.
• Initial and terminal properties attached directly to their owned span, preserving metadata across collection boundaries. Cross-thread regression controls preserve lazy disabled behavior.

Intervals overlap. An owner dropping does not establish physical memory reclamation, and an exporter flush is not a delivery acknowledgement.

Actual verification

Exact commit 98de82ab0f58385c795b33ec1dc0db85a362a998, tree dc91076ecbb87d183272a56759c0ab773883b523, passed the candidate and clean committed native macOS x86_64 gates: versions, formatting, Clippy with warnings denied, all workspace tests, the separate 35-test Iceberg descriptor control, and optimized release compilation. All nine steps and the original inner/outer waits returned zero; independent current source, binary, process and lock closure passed.

Release: Rust 1.97.1, optimization 3, full LTO, one codegen unit, debug information disabled, stripping enabled. Binary: 150,955,636 bytes, SHA256 a4842a8c2133738d12a20cc2322b4bcd440e8a93a45c842a12987079ae04071c.

The repaired smoke and full main05 campaign passed strict received metadata and causal qualification: 60 cold/warm pairs, 120 complete 4,096-group answers, 20 pairs at each P4/P16/P32, and both actual workers in every action. Separate typed-error, pool-refusal and cancellation controls passed. All 43 named phases are observed across the success and conditional fault controls; operation bindings are complete. Original waits returned zero, no forced cleanup occurred, and current recorded processes/groups and all three locks were absent at the final closure.

Queryable phase and sampled-resource summaries are retained. Sampling identity gaps and absent memory/buffer gauges remain explicit; this does not establish PSS, unique physical memory, an OS 32 GiB envelope or native-pool fit.

The original source85/main03 failure remains retained: its sixth pair had correct answers but two received stream spans lacked initial metadata. The dependency control reproduces the collection/property ordering defect, and the repaired current campaign passes without weakening delivery, causal, answer or conditional-phase predicates. Earlier failed source, delivery and metadata-closure attempts remain distinct evidence.

The Sem completion report is being updated with these closed results and the separate two-host controls. This draft makes no historical stream-failure diagnosis or whole-process physical-memory claim.