fix: heartbeat default kills live agents; 797th manifest undoes production on re-apply

[Nideesh1]

Two related failures, both of which stop a collector agent from ever completing a run.

1. The heartbeat default kills agents mid-turn

PR #17 raised heartbeat::DEFAULT_UNRESPONSIVE_TIMEOUT_SECS to 600 — but production never reads it. kernel::start_heartbeat_monitor (kernel.rs:4865) builds its config as:

HeartbeatConfig {
    default_timeout_secs: self.config.heartbeat.default_timeout_secs,
    ..HeartbeatConfig::default()
}

The [heartbeat] TOML value overrides the constant. With no [heartbeat] section present it falls back to default_heartbeat_timeout() = 180, so #17 is dead code on every real deployment.

Observed on a live collector, where a turn is allowed 600s (DEFAULT_AGENT_TURN_TIMEOUT_SECS):

16:47:18  shell_exec ok
16:51:02  Unresponsive Running agent marked as Crashed, inactive_secs=207
16:51:32  Auto-recovering crashed agent (attempt 1/3)
16:52:02  Agent recovered successfully

Killed at 207s, in-flight work discarded, roughly three times per allowed turn — forever. An agent waiting on a slow or rate-limited model could never finish one.

• default_heartbeat_timeout() 180 → 900
• doc comment recording that this is the value production uses, not the heartbeat.rs constant
• new test_heartbeat_default_exceeds_agent_turn_timeout asserts the invariant rather than the literal, so the two cannot silently drift apart again

2. The 797th manifest would undo production if re-applied

deploy/797th/openfang.yaml had drifted far enough that applying it — the normal way to change anything in it — would break the deployment it describes:

	manifest had	live cluster
default_model	stepfun-ai/step-3.5-flash — 410 Gone since 2026-07-27	thinkingmachines/inkling
[knowledge]	absent	11 entity types, 15 relations
[heartbeat]	absent	needed, see above

The dead model pin is the worst of the three: it is the global default, so restoring it takes down every agent at once. That already happened once. The [knowledge] block is the vocabulary the kernel enforces on the graph — without it collectors silently revert to inventing one-off labels.

falkor-sync.yaml was also missing has_new_artifacts() and its call site — the guard that keeps a quiet cycle from delivering an empty "no new artifacts" report every run. Verified the repo copy now matches the deployed script byte-for-byte apart from comment wording.

3. Collector briefs are now in the repo

Each collector instance is steered by a context.md re-read every turn (agent_context.rs::load_context_md) — no rebuild, no restart. Those files are the only definition of each beat and existed nowhere but the PVC; a lost volume meant rewriting them from memory. Added deploy/797th/collector-briefs/ with the underground and remote-jobs briefs and a README covering the report-format contract the delivery job depends on.

Verification

• CI: build / test / clippy / format green on all three platforms
• Security Audit fails on RUSTSEC-2026-0222 (wasmtime) — a newly published advisory, not introduced here. This branch changes no Cargo.toml or Cargo.lock; the diff is one Rust constant plus YAML and Markdown
• YAML + TOML of both manifests parsed and asserted against the live cluster values
• sync.py extracted from the manifest and py_compiled
• Not compiled locally: the only available host is a k8s control-plane node where cargo has previously OOM-killed kube-apiserver

Applying without waiting for merge

[heartbeat]
default_timeout_secs = 900

🤖 Generated with Claude Code