fix(swarm): make the multi-persona campaign actually work (stdout was corrupting ACP)

[New1Direction]

What

The campaign's "theatrical swarm" was a stdout-pollution protocol bug, not fake work. Every korg worker subprocess was doing real work (git worktree, applied patch, real measured diff, signed results) and succeeding — but the leader recorded each one as crashed (exit_code=-1).

Root cause: the leader parses each worker's stdout as newline-delimited ACP/JSON. Log output was going to stdout too, so the worker's first reply was preceded by a tracing line (2026-…) → the leader's first read_acp_envelope failed to parse → EOF → false crash, before reading any result. This hit every persona in both deterministic and ollama campaigns. SP1–SP4 fixed the in-process paths; this was the subprocess path, invisible until a real campaign's worker stdout was read.

The fix (stdout discipline)

• korg-core/telemetry.rs: init_tracing had no .with_writer → defaulted to stdout (while ironically checking stderr for ANSI). Logs now go to stderr.
• korg-runtime/harness.rs: a stray [TelemetryEmitter] println! → eprintln!; the legacy run() path's println!s → eprintln! (same latent hazard).

The worker already sent the authoritative TerminationReport (real success/doom_loop status + terminal_tx_id) — it was just never readable. (An earlier extra report I added was a misdiagnosis; review caught the double-send and it was removed. The e2e still passes, proving the stdout fix alone is the cure.)

Campaign provider selection + live reliability

• src/main.rs: global --provider/--model/--base-url, exported as env at startup so every worker subprocess builds the selected provider via KorgConfig::load() (no config threading; covers TUI/web). run-once unified onto these flags (no behavior change: no flag → deterministic).
• korg-runtime/session.rs: worker spawn forwards the LLM env to the child.
• korg-runtime/personas.rs: implementer personas (Benjamin/Lucas) request response_format: json_object so live models reliably emit a parseable mutations block; prose personas unchanged; deterministic stub ignores it.

Proven

• Deterministic campaign: all 4 workers terminate success, Benjamin attests a real measured mutations=1, DAG data-flow real (Benjamin's payload carries Captain+Harper output; Lucas's carries Benjamin's mutation).
• Live ollama campaign: all personas complete on a real model (~5 min real latency); an implementer produced a real mutations=1.

Test plan

• Gated e2e tests/campaign_e2e.rs (#[ignore]) — runs the real korg campaign, asserts workers complete, none crash, Benjamin attests mutations=1 (passes 71s locally)
• cargo test -p korg-runtime — 143 pass; cargo test -p korg-core pass
• fmt + clippy(correctness) clean
• Independent Rust review (caught + fixed the double-TerminationReport)
• CI green on the branch (timeout-guarded; the campaign e2e is #[ignore], not run in CI)

Also

• Renames non-standard Tests/ → tests/ so integration tests are discoverable cross-platform.

Honesty unchanged: the mutations count is the real git diff measurement; this only fixes whether a worker is recorded as done and which provider it uses.

Known pre-existing follow-ups (not introduced here)

• doom_loop exit status maps to crashed=true (session.rs:467 / workers.rs:870) — a controlled doom-loop signal shouldn't be a crash.
• The post-work read_acp_envelope in run_as_stdio_worker has no timeout (works in practice — the leader drops stdin → EOF).