Backfill prompt sessions for origin filters

[tony]

Summary

• Keep prompt-scope searches on the prompt-history fast path for bare text queries.
• Backfill prompt-capable session/conversation sources when prompt-scope searches include origin filters such as cwd, repo, worktree, or project fields.
• Preserve prompt result filtering while allowing metadata-bearing session stores to participate.

Verification

• rm -rf docs/_build; uv run ruff check . --fix --show-fixes; uv run ruff format .; uv run ty check; uv run py.test --reruns 0 -vvv; just build-docs;

Goal

Idempotent completion condition for /goal. It asserts an end state, not an action: if the state already holds, the first evaluation passes and nothing is edited.

/goal Work only in a worktree on PR #98's branch fix/origin-filter-session-prompts at the pushed PR head; its existing worktree is 249 commits stale, so refresh it. Never touch the shared master tree, never push. First report git rev-parse --abbrev-ref HEAD, git rev-list --count HEAD..master (23 today), rg -c origin_fields src/agentgrep/query/compile.py (2 here, 0 on master) and ls tests/test_origin_prompt_sources.py (absent today). If they read that branch, 0, 2 and a module that exists and passes, make no edits and stop. End state there: rebased onto local master, with tests/test_origin_prompt_sources.py present and green under .venv/bin/python -m pytest, proving that a prompt-scope search with an origin constraint (cwd, repo, worktree or branch; as an explicit filter and as a compiled query field) surfaces prompts from the origin-bearing session stores of an agent that also has a prompt-history store, while the same query without one still takes the prompt-history-only path and opens no session store. Constraints: keep the five-file engine change in shape, no unrelated refactor; prompt scope still emits only prompt-kind records; CompiledQuery.origin_fields stays planner-internal - no facade re-export, frontend consumption or public-API docs. Do not frame this as an implicit deep search and add no also-search-conversations flag: scope keeps selecting record kinds while a later effort axis selects how much source material is read. PR #72 also extends CompiledQuery there; do not pre-resolve that conflict. Finally report exiting 0: rm -rf docs/_build; uv run ruff check . --fix; uv run ty check; NO_COLOR=1 env -u VIRTUAL_ENV -u UV_NO_SYNC .venv/bin/python -m pytest -m "" --reruns 0 -q; just build-docs. Or stop after 12 turns and report what remains.

• Proof — git rev-parse --abbrev-ref HEAD; git rev-list --count HEAD..master; rg -c origin_fields src/agentgrep/query/compile.py; NO_COLOR=1 env -u VIRTUAL_ENV -u UV_NO_SYNC .venv/bin/python -m pytest tests/test_origin_prompt_sources.py -q; rm -rf docs/_build; uv run ruff check . --fix; uv run ty check; NO_COLOR=1 env -u VIRTUAL_ENV -u UV_NO_SYNC .venv/bin/python -m pytest -m "" --reruns 0 -q; just build-docs
• Idempotence — Once the worktree HEAD is the branch, its behind-count against local master is 0 and the regression module exists and passes, the probes report satisfied on the next turn and no rebase or edit happens.
• Blocked by — nothing - the branch merges into master today; its only overlap is Add OpenTelemetry local observability #72, which also extends CompiledQuery in query/compile.py and is explicitly left unreconciled here
• Blocks — nothing
• Preserves — scope stays a record-kind selector and is never fused with effort, so a later progressive-depth ladder still owns how much source material is read; CompiledQuery.origin_fields stays planner-internal so predicate placement can move without a public-field break; the Add OpenTelemetry local observability #72 field-predicate overlap is left for its own reconciliation