Node state-transition model: execution_id stamping + derived statuses

[Pino de Candia (pinodeca)]

Summary

Implements read-time node-status inference for pg_durable so that
df.instance_nodes() and df.explain() report structure-aware statuses for each
node, including branches that were never taken and nodes superseded by later loop
iterations.

The implementation keeps the physical node statuses as pending, running,
completed, and failed. Derived statuses such as skipped, and loop
re-entry-as-pending, are computed from graph structure plus each node's
status_details.execution_id lineage.

What changed

Schema

• Adds a nullable df.nodes.status_details JSONB column holding the worker-written
  {"execution_id": "<instance::generation[::...]>"} stamp. The column is added
  for fresh installs in src/lib.rs and for existing installs through the
  0.2.3 -> 0.2.4 upgrade script.

Write path

• Stamps each running and terminal node transition with deterministic execution
  lineage.
• Applies a monotonic generation fence so stale updates from superseded
  older-generation executions cannot overwrite newer node state.
• Keeps the new .so compatible with older schemas that do not yet have
  df.nodes.status_details by probing for the column at runtime.

Read path

• df.instance_nodes(instance_id) returns one row per node with stored physical
  status plus status_details, inferred_status, and
  inferred_status_from_ancestor_id.
• inferred_status is derived by walking the graph from the root:
  • a non-terminal node under a terminal ancestor reads as skipped;
  • a node from a prior loop iteration reads as pending because it will re-run;
  • a node that physically ran in the current generation keeps its stored status.
• df.explain(instance_id) uses the same inference helper so the tree view agrees
  with df.instance_nodes().
• df.instance_nodes(text, integer) remains callable for old schemas and upgraded
  callers as a shape-compatible adapter. It no longer queries execution history,
  ignores last_n_executions, and returns one row per node with execution_id = 1.

Tests & docs

• Adds tests/e2e/sql/53_inferred_status.sql covering untaken df.if() arms,
  failed df.then() downstream nodes, clean sequences, and df.explain() output.
• Updates USER_GUIDE.md, docs/api-reference.md, and the pg-durable-sql skill
  for the inferred-status model and the new primary df.instance_nodes() shape.

Related issues

• df.instance_nodes leaves race-loser nodes running or pending after race completion #171: race-loser nodes are addressed from a reporting perspective. The
  abandoned RACE arm now reads as skipped via inferred_status; the physical
  df.nodes.status remains unchanged.
• No "skipped" indication for downstream steps after a step fails #240: the right side of a failed df.then() now reads as skipped.

Upgrade & migration

• The 0.2.3 -> 0.2.4 script adds df.nodes.status_details and recreates
  df.instance_nodes() so upgraded schemas match fresh installs.
• The new binary remains compatible with previous 0.2.x schemas before
  ALTER EXTENSION UPDATE: write/read paths avoid referencing status_details
  when the column is absent, and the old two-argument df.instance_nodes() symbol
  remains callable.
• Drain in-flight instances before upgrading. The worker now records execution
  stamps in activity inputs, and duroxide compares recorded activity inputs by
  exact equality during replay.