feat(vigil): docs and e2e fixtures (slice 5 of 5)

[allen-munsch]

Summary

Slice 5 of 5 in the vigil decomposition: documentation and the e2e fixture harness. No source changes — this slice is docs and test scaffolding only, so the CI matrix is unchanged from slice 4.

What's in

• docs/proposals/vigil.md — the feature proposal (trigger taxonomy, rite gate, observance, the toll/watcher/harbinger trigger types).
• docs/vigils/* — usage, use-cases, and a loop/MCP comparison.
• tests/fixtures/vigil/* — podman-compose harness with Jenkins/Prefect/Airflow echo services, poller plugins, sanity-check JSON, setup scripts, and a watch-dir trigger fixture.

What's deliberately not here

The feature/vigil-phase5 branch this was ported from sits on pre-0.24.1 main, so its remaining diff against slice 4 is dominated by version drift that is not vigil work and is correctly excluded here:

• /effort reasoning-effort support (src/config/mod.rs, src/cli.rs, src/plugin/mod.rs, src/agent/agent_loop/*).
• ThinkingLevel::Max and the max_tokens budget fix.
• The ServiceError boxing in src/provider/adapter.rs.
• Cargo.toml/Cargo.lock/nix/bin.nix/CHANGELOG.md version bump (0.24.1 vs 0.24.0).

Two vigil-adjacent divergences from phase5 are also left out, both worth a deliberate decision rather than slipping into a docs PR:

1. vigil is not in default features. phase5 adds 'vigil' to the default feature set; slices 1–4 kept it opt-in (vigil = ["dep:notify"]). Flipping it would pull notify into every default build and change what windows-default compiles, so it needs a separate call.
2. Obsolete vigil tests remain. phase5 deletes src/tests/vigil_tests.rs and the test module in src/extras/vigil_db.rs (superseded by the keeper-level tests in src/extras/vigil/tests.rs); slices 1–4 still carry them. A follow-up cleanup PR should remove them.

Verification

• cargo fmt --all --check — clean.
• All 26 files are new (no Rust, no tracked-file modification); the only working-tree untracked files are the two pre-existing DAP fixture binaries, left unstaged.
• Fixture plugins reference only symbols that exist after slices 1–4: harness/json-decode, harness/notify, harness/register-command, vigil/emit.

[allen-munsch]

Quick note on the one-line evil-janet change, since the vendored diff is large and mostly just the vendored source:

The bug: Janet's garbage collector never marks janet_vm.top_dyns — the table that holds dynamic bindings when no fiber is active. dirge hits this because it calls janet_setdyn outside any fiber, so those bindings land in top_dyns. On the next GC the table gets swept, and a later janet_dyn (via the error printer) dereferences freed memory — the SIGSEGV.

The fix is one line in vendor/evil-janet/csrc/janet.c, in janet_collect, added right after the root fiber is marked:

if (janet_vm.top_dyns) janet_mark_table(janet_vm.top_dyns);

That tells the collector to treat the table as a GC root, exactly how every other VM-global GC pointer in that function is already handled. Upstream Janet (1.37.2 through 1.41.3 and current master) has the same missing root, which is why this had to be a vendored patch rather than a version bump.

edit: I'm not sure if this should be vendored, or if we should attempt a PR to upstream