[Feat]: Add a task-state-invariants agent skill for sibling-surface propagation #1413

Is your feature request related to a problem? Please describe

The agent rule layer in this repo is good at three things: code style
(docs/ai/coding_conventions.md), the pre-done gate
(docs/ai/mandatory_checks.md), and post-mortem capture
(.agents/skills/mistake-reflection). What none of them state is which
surfaces a task-state write has to reach. That gap keeps producing the same
class of defect: a value that is correct on one surface and missing or stale
on another.

Recent instances, all in this class:

• v0.3 gRPC compat drops task status timestamp and three other fields #1276 (closed): status.timestamp was mapped by
  src/a2a/compat/v0_3/conversions.py (JSON-RPC/REST compat) but not by
  src/a2a/compat/v0_3/proto_utils.py (gRPC compat) — the same field appeared
  or disappeared depending on transport.
• [Bug]: Framework-written FAILED status drops status.timestamp, persisting last_updated as NULL #1331 (open): a framework-written FAILED status carried no
  status.timestamp, and src/a2a/server/tasks/database_task_store.py:152-157
  derives last_updated from it, so the row persisted NULL and dropped out
  of ListTasks ordering.
• [Bug] DefaultRequestHandlerV2: ActiveTaskRegistry resumes from a stale task snapshot after input_required, silently overwriting store state (multi-replica deployments) #1188 (open) / [Bug]: Blocking continuation returns stale INPUT_REQUIRED when executor emits current task #1409 (open): the handler resumes or returns from a snapshot
  that disagrees with what later reaches the store and the stream.
• [Bug]: taskId without contextId creates a new context instead of inheriting it #1206 (closed as stale) → [Bug]: taskId-only continuation still fails required contextId inference in SDK 1.3.0 (follow-up to #1206) #1411 (open): contextId inference makes events
  disagree with the stored Task.
• Review findings on fix(server): always stamp status.timestamp on terminal task states #1346: stamping the local copy after
  task.status.CopyFrom(event.status) leaves the published event unstamped
  (src/a2a/server/tasks/task_manager.py:285), and the CANCELED paths at
  src/a2a/server/agent_execution/active_task.py:971-973 keep the previous
  state's timestamp. These were caught by review, not by any documented rule.

The same confusion shows up in the discussion forum, where it has no issue
tracker to land in: #248 ("Should Message events terminate…", 6 comments),
#435, #40 ("What's the purpose of EventQueue?"), and the multi-replica thread
#1224 / #336.

Describe the solution you'd like

A new skill, .agents/skills/task-state-invariants/SKILL.md, in the same
shape as mistake-reflection (trigger list → procedure → anti-patterns),
covering:

1. the invariant set, each claim anchored to current main (terminal states
   need a timestamp; stamp the event before CopyFrom; CopyFrom inherits
   the old timestamp; push delivery is best-effort; zero events means a hung
   caller; store and stream must agree; v0.3 compat has two independent
   mappers; continuation identity is inferred per spec §3.4.3);
2. a sibling-surface checklist to run before declaring such a change done
   (persisted store, streaming subscribers, push notification, cluster
   task_events table, both compat mappers);
3. a test map naming the existing file per surface, so coverage lands where
   the surface is;
4. the evidence a PR in this area should carry, consistent with
   docs/ai/evidence_rules.md.

Plus a one-line pointer in AGENTS.md "Mandatory workflow". That pointer
matters: AGENTS.md is the only file in the repo that routes to
.agents/skills/, so a skill without a route is never loaded.

Describe alternatives you considered

• A docs/ai/ rules page. Rejected as the primary form: docs/ai/* is
  read at task start, while this knowledge is only useful at the moment you
  touch a status write — which is how mistake-reflection is structured.
  Happy to have both if maintainers prefer the rule text to live in docs/ai
  with the skill as a thin pointer.
• Encoding it only in tests. Partially useful, but tests cannot tell a
  change agent which sibling surface they forgot.
• Fixing each bug as it arrives. Six-plus in about six weeks suggests the
  per-bug loop is not converging.

Additional context

Scouting the rule layer, .agents/skills/ has exactly one skill today and it
was introduced internally by #1059 — I found no precedent for an externally
contributed skill, so raising this first rather than assuming. A draft
implementation (skill + the one-line router pointer, markdown-only, no
runtime code) is ready and filed as a PR linked to this issue. Easy to trim:
drop the AGENTS.md line, or scope the skill to the compat-mapper and
timestamp rules only.

Code of Conduct

• I agree to follow this project's Code of Conduct