Record five-pass song swarm run

.gitignore

@@ -6,5 +6,14 @@ TestResults/
 artifacts/
 .voidbot/
 target/
+.dotnet*/
+.nuget/
+.sandbox-home/
+.tmp/
+Microsoft/
+NuGet.Local.Config
+*.binlog
+*.log
+_tmp_*.txt
 *.user
 *.suo

state/evidence.jsonl

@@ -271,3 +271,4 @@
 {"timestamp":"2026-06-01T01:58:00+01:00","topic":"song-swarm-render-failure-visibility","claim":"Song-swarm render failure must be visible to agents and operators. `song-score` now writes `audio/render-failure.txt` and includes failure summaries in `evaluator-report.md`. The parser accepts documented `env=ad attack=... decay=...` syntax, because the manual taught that form and agents copied it. Codex musician jobs now set repo-local `DOTNET_CLI_HOME`/`NUGET_PACKAGES` with first-time-use disabled so internal renderer `dotnet run` calls do not write to denied sandbox home directories.","evidence":"src/AquaSynth.Core/PatchScript.cs; tests/AquaSynth.Dsl.Tests/PatchScriptTests.cs; tools/IpaTrialWorker/Program.cs; tools/run-song-snippet-trial-swarm.ps1; artifacts/parity/song-render-failure-smoke/20260601T0145-agent01-a01/smoke-agent01-target01-attempt01/song-candidates/agent_01__swarm__a01/audio/render-failure.txt; artifacts/parity/song-render-failure-smoke/20260601T0152-agent01-a01-after-ad/smoke-agent01-target01-attempt01/summary.csv"}
 {"timestamp":"2026-06-01T02:08:00+01:00","topic":"song-swarm-render-environment-and-syrinx-scaffold","claim":"Renderer builds must not depend on user-profile NuGet or dotnet config. `FaustCompiler.RenderAsync` now runs its temporary C# render project with temp-local HOME, APPDATA, LOCALAPPDATA, DOTNET_CLI_HOME, and NUGET_PACKAGES. The music manual and swarm prompt now treat acoustic/syrinx as complete-graph syntax: agents should start from known-valid syrinx patches or log a gap, not submit isolated `source_port`/`acoustic_network` fragments that cannot render.","evidence":"src/AquaSynth.Faust/FaustCompiler.cs; docs/music-production-knowledge/aqua-dsp-usage-manual.md; tools/run-song-snippet-trial-swarm.ps1; artifacts/parity/song-snippet-swarms/20260601T005553312/pass-001/agent-01/iterations/attempt-01/target-01/agent-01-target-01-attempt-01/song-candidates/agent_01__swarm__scene_a/audio/render-failure.txt"}
 {"timestamp":"2026-06-01T02:33:00+01:00","topic":"learned-only-music-knowledge-show-standard","claim":"A cumulative music-production knowledge loop should not fail when `music-distill --learned-only true` omits the seed `music-production-quality-standard-v1` document. The pass-end standard snapshot is display-only: it may show the canonical seed record when present, but if learned-only output only contains current curator/manual/failure/abstraction documents it must write an index-backed pass snapshot and continue the swarm.","evidence":"tools/run-song-snippet-trial-swarm.ps1; artifacts/parity/song-snippet-swarms/20260601T010706781/launcher stderr showed `No music knowledge document found for music-production-quality-standard-v1` after pass-001 despite successful learned-only distillation"}
+{"timestamp":"2026-06-01T04:23:00+01:00","topic":"five-pass-song-swarm-composition-run","claim":"A 5-pass, 2-agent, 2-task, 30-second song swarm over the Reinier hierarchy completed end-to-end with post-pass curator/distill/index promotion after every pass. The cumulative learned-only music knowledge loop stayed alive through all five passes. Official candidates improved from pass-001 top score 0.432376 to pass-004 top score 0.471762 and pass-005 top score 0.468977, with low mode-collapse risk around 0.07 and motion coverage around 0.638. Remaining pressure is producer discipline and syntax hygiene: official verdicts were mostly `pressure`, with two official render failures and one weak mode-collapse; one failure came from agent-authored newline loss gluing `soft_clip=true` to `meter` as `truemeter`.","evidence":"artifacts/parity/song-snippet-swarms/20260601T012948062; pass-005/top-scoring-candidates.md; pass-004/top-scoring-candidates.md; pass-*/knowledge-curator/knowledge-curation.md; pass-*/music-production-knowledge-report.md"}