[Guidance78]

The problem I hit

Ringer runs tasks as a flat set — everything is eligible at once. I run producer
tasks followed by an independent QC reviewer on a different model family, and
there's no way to say "review this one after it builds."

So every stage becomes two runs with me in the middle, doing nothing but
watching for the first one to finish and then launching the second. The
expensive part isn't the worker tokens, it's the directing model reloading
context to do dispatch work it could have declared up front.

The change

An optional task field:

{"key": "schema-qc", "depends_on": ["schema-producer"]}

• Tasks without depends_on behave exactly as they do today.
• A dependent waits before acquiring a max_parallel slot, never while
  holding one, so a two-stage manifest at max_parallel: 1 can't deadlock.
• It runs only after every prerequisite reaches a final PASS — a producer that
  fails attempt 1 and passes attempt 2 has passed.
• If a prerequisite doesn't pass, the dependent is skipped: no worker
  launched, no attempt consumed, and no eval row for a model that never ran.
  Skips propagate down the chain.
• Waiting, skipped and blocked_by show up truthfully in run state, the
  progress bar, the briefings and the summary.
• Missing keys, self-dependencies, duplicates, non-string entries, bare strings,
  dicts and cycles are all rejected at parse time, with the cycle path named.

Every task publishes its terminal result exactly once from a finally, so
preparation failures, timeouts, exhausted attempts, cancellation and unexpected
exceptions release waiters instead of hanging the run.

Deliberately not in this PR

No cross-run dependencies, no conditional workflow language, no dynamic task
generation, no automatic artifact routing or patch application. depends_on is
a control dependency only. Because a passing task's worktree can be removed
before its dependent starts, anything the dependent needs has to be exported by
the producer's check to a durable path outside the worktree — documented beside
the field.

Proof

• 27 new tests in tests/test_dependencies.py, all observed failing before the
  implementation existed.
• Disabling the dependency wait turns 7 of them red, including the fan-in
  ordering test — they detect the feature's absence rather than asserting on end
  status.
• Full suite: RINGER_NO_SELF_UPDATE=1 python3 -m unittest discover -s tests —
  280 pass on Python 3.12.
• git diff --check clean.
• Also exercised end to end on a real run: a producer that failed left its
  dependent skipped with 0 attempts, 0 tokens and no eval row, while a
  separate passing producer released its reviewer normally.

[Guidance78]

Rebased this PR onto the latest upstream main (2026-10-05). The staged-dependency regression suite passes (27 tests), and GitHub now reports the PR mergeable. The feature remains a single commit.