Funnel host-effect traces through gunbc.output_policy (kill the [file] read floor spam)

[briansrls]

Follow-up slice to #5874 (the deferred host-trace funnel).

Problem

[file] read / [file] write / [rest] / [hermetic:mock] were unconditional
eprintln!s. [file] read alone floods the CI floor log with hundreds of
source-file reads (doubled by the within-run double-resolve) — the spam the original
"CI output is a horror" report called out.

Fix — consume the .dag output policy (no Rust re-derivation)

• gunbc.output_policy gains ChannelPolicy + resolve_channel_policy(verbose, quiet):
  it enumerates the channels and calls channel_decision once each, so the seed
  evaluates the whole policy in one .dag call. channel_decision stays the single
  authority for channel × verbosity → decision.
• cli_run::install_output_policy evaluates resolve_channel_policy from the .dag
  authority at the current cli_verbosity and installs the per-channel verdicts into a
  process-global (set once before discovery threads spawn). It only transports the
  evaluated OutputDecision variants across the boundary — the policy logic is not
  mirrored in Rust.
• v1_interpreter routes the four trace sites through trace_emit(channel, line):
  reads/mocks → Instrumentation (Verbose-only), writes/rest → ShellTrace (condensed
  at Normal). render_shell_trace now consumes output_decision(ShellTrace) too,
  dissolving the cli_verbosity mirror terminal output modeling for CI/interactive #5860 left in it.
• claim_executor + claim_batch install the policy at startup. Unset (other bins)
  falls back to Full = pre-funnel behavior, so nothing else regresses.

Verified

A WET filesystem_read witness prints 0 [file] read lines at Normal and 1 at
Verbose, and still passes. output_policy_witness covers resolve_channel_policy
(Normal: Instrumentation=Suppressed / ShellTrace=Condensed / Diagnostic=Full; Verbose:
all Full). cargo fmt --all --check + clippy --all-targets clean.