feat(context): stateful compact serving for agent context tools

[edwin-zvs]

Problem

Input-token efficiency review of construct_context / agentd_context found five gaps (the etag work in #865 only landed on the MCP path):

1. Smith's native agentd_context bypassed all compaction — it pretty-printed the full AgentdContext on every call: 17-bullet memory policy + 13-bullet widget policy (~5KB), the full 12-entry Markdown-extension registry (~6KB), full memory files, and the full program-run payload. Roughly 3k tokens of static text per call before any real content, on the harness that calls the tool most.
2. Monolithic program-run etag — computed over the whole struct, so any program edit (including the mandatory first status-only shimmer edit) resent ~4KB of static run instructions + smart-clip reference along with the markdown.
3. Memory resent after the agent's own edits — the agent can't compute the server-side FNV etag, so right after writing a memory file it got the full content back.
4. Dedup relied on the model echoing known_* etags — unenforced and unmeasured.
5. Static repeats — instructions, session_widgets, and the reference hint were resent on every call.

Change

One shared compact-serving layer in construct_protocol::agent_context, used by both construct-mcp and smith's native tool:

• Per-process serve state: the serving process (one process = one agent) remembers what it sent. Repeat calls collapse unchanged memory/program payloads to {path, etag, unchanged} markers and send static fields once. known_* echoes are still honored (covers a restarted serving process mid-conversation).
• Split program-run etags: markdown and the static run reference (instructions + smart clips) are tracked independently — a program edit resends only the markdown. Status-only shimmer edits don't change markdown, so the required planning pass now costs ~nothing on the next context call.
• skip_memory: true: omit memory content the agent already holds verbatim (e.g. it just wrote the files); paths + etags still returned, and the content is marked served.
• refresh: true — the compaction escape hatch: resends everything. Everything omitted stays disk-recoverable (memory by path, program via construct_program_get), so a stale serve state degrades to an extra fetch, never data loss.
• Smith resets serve state natively on auto-compact, manual /compact, and destructive prune — the exact moments the model loses served content — closing the compaction/etag hazard on the native side entirely.
• Dropped known_* from the tool schema (still accepted), trimmed truncated/remaining_bytes to truncation cases only.

Not done (deliberate): version-based markdown diffs for program runs — partial markdown risks the agent forming a wrong mental model of the document; the etag split captures nearly all the win.

Spec 0095 updated with the serving contract (identical behavior across surfaces; native tools must not bypass the budget layer).

Testing

• New unit tests for the shared layer (first-serve/repeat-serve, refresh, split etags, skip_memory, known-etag echo, include_reference) in construct_protocol::agent_context.
• MCP dispatcher wiring test updated to exercise the process-level state.
• Full cargo test --workspace: all 19 targets green.

🤖 Generated with Claude Code