Live graph: tell a timeout apart from a 404, and retry only the timeout

[AaryanG21]

What this adds

The live planner could not express the difference between a failure worth
retrying and one that never will be. A failed node was terminal, and
apply_patch refuses to re-add an existing task id, so there was no vocabulary
for "that timeout deserves another attempt, that 404 does not".

That gap was not theoretical. On the shipped non-browser benchmark, 4 of 14
cases returned a completely empty answer: the planner attached the next stage
to a node that had failed, GraphStore.ready() only releases a node once every
parent has succeeded, so the child sat in pending forever, the executor ran
out of work and the run returned silently with nothing.

• core/live_graph/retry.py — failure classification and a bounded backoff
  policy. Classification is type-first: a known permanent exception type
  short-circuits before any message inspection, so an attacker-controlled path
  cannot smuggle the word "timeout" into a PermissionError and buy a retry of
  a refused sandbox escape.
• GraphPatch.retry — a retry creates a new node that inherits the failed
  node's parents and adopts its children. Nothing is edited in place: every
  attempt keeps its own state, its own error and its own place in the journal.
• GraphStore enforces the attempt ceiling and one-attempt-per-failure
  independently of the planner, so an LLM planner emitting retry cannot loop
  the budget away.
• The planner classifies before it plans around a failure, resolves its
  branches by retry lineage, and no longer parents the distiller on a failed
  node.

Benchmark, same 14 cases, before → after:
10/14 completed, 4 empty answers, 4 stranded nodes → 14/14 completed, 0
empty answers, 0 stranded nodes.

Proof

README.md carries the required section: capability, exact API request, graph
and ordered event trace, actual final result, evidence and provider/agent
assignments, the adversarial failure and its fix, and the commands to reproduce
from a fresh checkout with no provider key and no GPU.

scripts/repro_retry.py runs both scenarios end to end against a local origin
that answers 503 twice then 200, and a second that answers 404 forever.

Tests

76 passing (44 upstream, unchanged, plus 32 added), fully offline — no local
Ollama or provider key needed to run the suite.
tests/test_retry_policy_adversarial.py keeps the pre-fix message-matching
classifier as naive_classify, so the regression is executable: each attack
must genuinely fool the old implementation and be refused by the new one. It
also covers duplicate and replayed planner decisions, crash-recovery replay,
and a retry result arriving after cancellation.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FvXm9XSmn5E9fdXGypJKAZ