InlineZig burndown: observable pipeline runtime wiring

[cuzzo]

Non-low-risk InlineZig cluster from the opaque-Zig hardening pass.

Sites:

• src/backends/pipeline_host.rb:3075 obs_alloc
• src/backends/pipeline_host.rb:3081 obs_wg_init
• src/backends/pipeline_host.rb:3085 obs_set_completion
• src/backends/pipeline_host.rb:3197 obs_consumer_spawn
• src/backends/pipeline_host.rb:3389 obs_reduce_publish
• src/backends/pipeline_host.rb:3424 obs_distinct_publish

Risk:

• Observable lowering hides heap allocation, WaitGroup initialization, completion ownership transfer, spawned-fiber context allocation/transfer, CAS publish loops, and submit calls inside InlineZig snippets.

Acceptance:

• Replace with structural MIR nodes or dedicated observable runtime operation nodes whose fields expose allocator, ownership transfer, spawn, and publish semantics.
• Preserve observable cleanup and WaitGroup ownership invariants.
• If a site remains InlineZig, document why a structural node increases complexity and why checker metadata is sufficient.

[cuzzo]

Progress pushed in 427d630 (Structuralize observable InlineZig wiring). Resolved observable InlineZig reasons: obs_alloc, obs_wg_init, obs_set_completion, obs_distinct_publish. Added a final-MIR regression in spec/pipeline_legacy_matrix_spec.rb so those structuralized observable reasons cannot reappear.

Remaining #102 sites:

• obs_consumer_spawn: still a critical multi-statement spawn scaffold that allocates/transfers the consumer context. Clean removal needs a dedicated observable-consumer spawn MIR/runtime bridge.
• obs_reduce_publish: CAS-loop publish is still emitted transiently into the spawn body; removing it cleanly likely needs a typed AtomicReduce publish node or runtime helper that can carry the reducer body safely.