Removes `knowledge_retrieval`, `pre_code_execution`, and `mcp_connect`
from the catalog — no concrete use case justifies them yet, and the mcp
seam had an observability gap (an abort fires before the connection
events, leaving no native record). The catalog now adds only `pre_step`
and `post_step` on top of the model/tool and execution-boundary points.
Pulls `tool_selection`, `memory_write`, and `memory_read` out of the
catalog. The tool seam needs a defined matching contract first (platform
and MCP tools expose different name shapes and attributes), and the
memory seams need real skip semantics instead of surfacing a policy
abort as an error log. Both are tracked as follow-up cards; the points
will return once those contracts are settled.
Removes `pre_delegation`, `retry_attempt`, `flow_transition`, and
`router_decision` — none has a hard use case yet, and `retry_attempt`
was wired at three unrelated seams (agent task retry, tool parse retry,
guardrail retry) with a different payload at each, so its semantics were
ambiguous. Also casts the `post_step` payload back to `TaskOutput` in
`task.py` so the seam keeps mypy green.
Wires the step, agent, subsystem, and flow points onto the dispatcher:
`pre_step`/`post_step`, `tool_selection`, `pre_delegation`, `retry_attempt`,
`memory_write`/`memory_read`, `knowledge_retrieval`, `pre_code_execution`,
`mcp_connect`, `flow_transition`, and `router_decision`, extending the typed-
context module with their contexts. Each seam passes a typed context whose
`payload` a hook may observe, mutate, or replace. Adds the conformance suite
for these points and a new `interception-hooks` doc page with the full
point/payload catalog and the interceptor contract.
