feat: compose full pipeline and LoCoMo system benchmark

.env.example

@@ -11,6 +11,18 @@ REMEMBERSTACK_SELFHOST_DEPLOYMENT_SLUG=local
 REMEMBERSTACK_SELFHOST_DEPLOYMENT_NAME=Local memory
 REMEMBERSTACK_SELFHOST_API_PORT=8000
 
-# Required to start the provider adapter. The short Markdown smoke path makes no
-# provider request; replace this value before processing a real corpus.
+# Required to start the provider adapter. Replace this value before processing
+# a corpus; the complete Compose pipeline makes extraction and embedding calls.
 REMEMBERSTACK_OPENROUTER_API_KEY=replace-before-real-use
+# Optional exact OpenRouter embedding-provider slug; unset keeps automatic routing.
+# REMEMBERSTACK_OPENROUTER_EMBEDDING_PROVIDER=nebius
+# Optional global reasoning effort for every chat-generation call. Allowed:
+# none, minimal, low, medium, high, xhigh, max. Unset uses the model default;
+# "none" reduces extraction latency but can reduce adjudication quality, and
+# models that require reasoning may reject it.
+# REMEMBERSTACK_OPENROUTER_REASONING_EFFORT=none
+
+# Optional benchmark/deployment model overrides. Keep explicit model IDs for a
+# reproducible run; do not use a rotating router such as openrouter/free.
+# REMEMBERSTACK_E2_EXTRACT_MODEL=nvidia/nemotron-3-super-120b-a12b:free
+# REMEMBERSTACK_E3_NORMALIZE_MODEL=nvidia/nemotron-3-super-120b-a12b:free

.github/workflows/ci.yml

@@ -53,28 +53,60 @@ jobs:
               raise RuntimeError("MinIO replaced an immutable object key")
           PY
 
-      - name: Ingest and observe the E0 chain
+      - name: Prove the zero-cost full pipeline
         run: |
-          curl --fail --data-binary $'# Hello\n\nCompose smoke.\n' \
-            'http://localhost:8000/ingest?filename=smoke.md&mime=text%2Fmarkdown'
+          running_services="$(
+            docker compose --env-file .env.example ps --status running --services
+          )"
+          for service in \
+            api worker-convert worker-structure worker-chunk worker-embed-chunk \
+            worker-extract-claims worker-normalize-relations \
+            worker-adjudicate-supersession worker-embed-claim worker-reconcile \
+            worker-label-relation; do
+            printf '%s\n' "$running_services" | grep -qx "$service"
+          done
+          ingested="$(
+            curl --fail --header 'Content-Type: application/octet-stream' \
+              --data-binary ' ' \
+              'http://localhost:8000/ingest?filename=empty.md&mime=text%2Fmarkdown'
+          )"
+          version_id="$(
+            printf '%s' "$ingested" |
+              docker compose --env-file .env.example exec -T api \
+                python -c 'import json, sys; print(json.load(sys.stdin)["version_id"])'
+          )"
           observed=''
-          for attempt in {1..30}; do
+          for attempt in {1..60}; do
             observed="$(docker compose --env-file .env.example exec -T postgres \
               psql -U rememberstack -d rememberstack -Atc \
-              "SELECT count(*) FILTER (WHERE stage='convert' AND status='succeeded'), count(*) FILTER (WHERE stage='structure' AND status='succeeded'), count(*) FILTER (WHERE stage='chunk' AND status='pending') FROM processing_state")"
-            if [ "$observed" = '1|1|1' ]; then
+              "SELECT count(*), count(*) FILTER (WHERE status='succeeded'), count(DISTINCT stage) FROM processing_state WHERE target_id='$version_id'")"
+            if [ "$observed" = '10|10|10' ]; then
               break
             fi
             sleep 1
           done
-          test "$observed" = '1|1|1'
+          test "$observed" = '10|10|10'
           test "$(docker compose --env-file .env.example exec -T postgres \
-            psql -U rememberstack -d rememberstack -Atc 'SELECT count(*) FROM cost_ledger')" = '0'
+            psql -U rememberstack -d rememberstack -Atc \
+            'SELECT count(*) FROM cost_ledger')" = '0'
+          docker compose --env-file .env.example --profile operations \
+            run --rm projections
+          readiness="$(
+            curl --fail --header 'Content-Type: application/json' \
+              --data "[\"$version_id\"]" \
+              'http://localhost:8000/readiness?require_projections=true'
+          )"
+          printf '%s' "$readiness" |
+            docker compose --env-file .env.example exec -T api python -c \
+              'import json, sys; report = json.load(sys.stdin); assert report["ready"]; assert all(item["ready"] for item in report["projections"])'
 
       - name: Stage an upgrade from the prior schema
         run: |
           docker compose --env-file .env.example stop \
-            api worker-convert worker-structure
+            api worker-convert worker-structure worker-chunk worker-embed-chunk \
+            worker-extract-claims worker-normalize-relations \
+            worker-adjudicate-supersession worker-embed-claim worker-reconcile \
+            worker-label-relation
           docker compose --env-file .env.example run --rm --no-deps \
             --entrypoint alembic setup downgrade p7_02_0016
           test "$(docker compose --env-file .env.example exec -T postgres \
@@ -116,7 +148,7 @@ jobs:
             docker compose "${compose_args[@]}" ps --status running --services
           )"
           test -z "$(printf '%s\n' "$running_services" \
-            | grep -E '^(api|worker-convert|worker-structure)$' || true)"
+            | grep -E '^(api|worker-(convert|structure|chunk|embed-chunk|extract-claims|normalize-relations|adjudicate-supersession|embed-claim|reconcile|label-relation))$' || true)"
           test "$(docker compose "${compose_args[@]}" exec -T postgres \
             psql -U rememberstack -d rememberstack -Atc \
             'SELECT version_num FROM alembic_version')" = 'p7_02_0016'
@@ -133,7 +165,11 @@ jobs:
           running_services="$(
             docker compose "${compose_args[@]}" ps --status running --services
           )"
-          for service in api worker-convert worker-structure; do
+          for service in \
+            api worker-convert worker-structure worker-chunk worker-embed-chunk \
+            worker-extract-claims worker-normalize-relations \
+            worker-adjudicate-supersession worker-embed-claim worker-reconcile \
+            worker-label-relation; do
             printf '%s\n' "$running_services" | grep -qx "$service"
           done
           curl --fail http://localhost:8000/healthz

Dockerfile

@@ -13,7 +13,8 @@ WORKDIR /app
 COPY pyproject.toml uv.lock README.md LICENSE alembic.ini ./
 
 RUN addgroup --system app \
-    && adduser --system --ingroup app app \
+    && adduser --system --ingroup app \
+        --home /var/lib/rememberstack --no-create-home app \
     && mkdir -p /var/lib/rememberstack/forget-manifests \
     && chown -R app:app /var/lib/rememberstack \
     && uv sync --locked --no-dev --extra server --no-install-project

README.md

@@ -11,10 +11,11 @@ at a million documents.
 > **Pre-release software.** Phases 0–7 are implemented and tested. The public
 > [`v0.1.0`](https://github.com/writeitai/remember-stack/releases/tag/v0.1.0) release is available
 > from PyPI and GHCR; release automation, trusted publishing, tag protection, and bounded
-> contributor-agreement enforcement are in place. The fresh-deployment Docker Compose skeleton
+> contributor-agreement enforcement are in place. The fresh-deployment Docker Compose profile
 > is documented under
-> [Self-host deployment](website/src/app/docs/deployment/page.mdx); it proves PostgreSQL, MinIO,
-> API ingestion, and the first two E0 worker stages, not a production rollout. The build follows
+> [Self-host deployment](website/src/app/docs/deployment/page.mdx); it composes PostgreSQL,
+> MinIO, the API, all ten continuous E/P1 routes, and explicit P2/P3 publication for one
+> deployment. It remains pre-release software, not a production rollout. The build follows
 > [plan/plans/roadmap.md](plan/plans/roadmap.md).
 
 ## TL;DR

benchmarks/locomo/README.md

@@ -1,17 +1,16 @@
-# RS-LoCoMo-v1 setup
+# RS-LoCoMo-Full-v1 setup
 
-This directory implements the reviewed `RS-LoCoMo-v1 J@30` protocol. It does not contain the
-LoCoMo data and does not auto-download it. The data is CC BY-NC 4.0; confirm the intended use
-before obtaining it from the
-[official repository](https://github.com/snap-research/locomo/tree/3eb6f2c585f5e1699204e3c3bdf7adc5c28cb376/data).
+This directory contains the unshipped full-system LoCoMo adapter. It does not vendor or
+auto-download LoCoMo. Supply the exact pinned `locomo10.json` only after confirming its
+CC BY-NC 4.0 terms.
 
-Install the repository and the two deterministic-scorer dependencies:
+Install the repository plus deterministic scorer dependencies:
 
 ```bash
 uv sync --extra benchmark
 ```
 
-The safe first command is local:
+The safe first command is local and makes no API or model call:
 
 ```bash
 uv run --extra benchmark python -m benchmarks.locomo prepare \
@@ -20,27 +19,35 @@ uv run --extra benchmark python -m benchmarks.locomo prepare \
   --output .benchmark-runs/locomo-smoke
 ```
 
-`prepare` checks the exact pinned SHA, validates the committed eight-question smoke manifest,
-renders session Markdown, and writes a call/document plan. It does not contact RememberStack or
-an evaluator.
-
-Do not proceed to `ingest`, `answer`, or `judge` until the owner walkthrough in
-[`locomo_benchmark_design.md`](../../plan/designs/locomo_benchmark_design.md#12-pre-run-checklist).
-Those stages require `--execute` plus exact isolation/readiness acknowledgements and explicit
-document, question, call, and shared evaluator-cost ceilings.
-
-Question and call ceilings are run-absolute, not allowances for one sample invocation.
-`--max-questions` must therefore cover the complete prepared tier (8 for smoke, 200 for
-development, or 1,540 for publication), while reader and judge call ceilings include calls
-already checkpointed for earlier samples. The evaluator-cost ceiling likewise covers the shared
-reader-plus-judge ledger for the entire run.
-
-The local ledger records provider usage attached to successfully parsed, checkpointed calls.
-Provider-billed calls that fail before usable accounting is returned are not reconstructable, and
-a process death after a response but before its atomic checkpoint may repeat that one call.
-Provider/account hard limits are therefore the hard monetary boundary; the harness ceilings are
-additional fail-closed operational guards.
-
-The released Compose profile is not yet a valid target: it wires only `convert` and `structure`,
-not the complete claim-indexing path. Use of a complete isolated deployment is a pre-run
-prerequisite, not something this harness silently constructs.
+The harness validates the pinned bytes, renders session documents, and fingerprints the
+eight-question smoke plan. Do not run remote stages until reviewing
+[`locomo_benchmark_design.md`](../../plan/designs/locomo_benchmark_design.md).
+
+The stock Compose deployment now includes all ten continuous E/P1 workers. After ingesting one
+isolated conversation and waiting for them to settle, publish the aggregate projections once:
+
+```bash
+docker compose --profile operations run --rm projections
+```
+
+The `answer` command then calls the public readiness endpoint. It refuses to run unless every
+requested version completed the exact composed stage generations and both P2/P3 builds began
+after that work completed. It also refuses a changed public recipe catalog. There is no manual “index ready”
+acknowledgement.
+
+Readiness also records the API process's current non-secret model configuration for operator
+review. Those values are not processing-time provenance; freeze one Compose environment for the
+run and retain the provider/cost artifacts.
+
+The primary protocol uses a bounded answer agent over normal public recipes, not hard-coded claim
+search. Limits are run-absolute: allow up to nine agent calls per selected question and one judge
+call per answer. The shared evaluator-cost value is a reported-spend stop threshold: a completed
+call can cross it, is recorded, and stops the run. Use the provider account cap as the hard
+monetary boundary. If that leaves later questions unanswered, they remain visible as zero-scored
+missing records; resuming them requires an explicitly higher threshold.
+
+P3 is built and freshness-checked as part of the ordinary deployment, but the remote recipe
+agent has no filesystem mount. This protocol therefore does not attribute answer quality to P3
+navigation. A future mount-enabled protocol needs a new fingerprint and name.
+
+No real benchmark has been run as part of the setup implementation.

benchmarks/locomo/cli.py

@@ -1,4 +1,4 @@
-"""Command line for the deliberately staged RS-LoCoMo-v1 harness."""
+"""Command line for the deliberately staged full-system LoCoMo harness."""
 
 from __future__ import annotations
 
@@ -51,10 +51,9 @@ def main(argv: list[str] | None = None) -> int:
                     run_dir=args.run,
                     sample_id=args.sample,
                     max_questions=args.max_questions,
-                    max_reader_calls=args.max_reader_calls,
+                    max_agent_calls=args.max_agent_calls,
                     max_evaluator_cost_usd=args.max_evaluator_cost_usd,
                     execute=args.execute,
-                    index_ready_confirmation=args.confirm_index_ready,
                     client=client,
                     provider=provider,
                 )
@@ -89,7 +88,7 @@ def _provider() -> OpenRouterModelProvider:
 
 
 def _positive_decimal(value: str) -> Decimal:
-    """Parse one strictly positive CLI monetary ceiling."""
+    """Parse one strictly positive reported-spend stop threshold."""
     try:
         parsed = Decimal(value)
     except InvalidOperation as error:
@@ -104,7 +103,7 @@ def _parser() -> argparse.ArgumentParser:
     parser = argparse.ArgumentParser(
         prog="python -m benchmarks.locomo",
         description=(
-            "RS-LoCoMo-v1: prepare is local; ingest/answer/judge require "
+            "RS-LoCoMo-Full-v1: prepare is local; ingest/answer/judge require "
             "explicit execution acknowledgements"
         ),
     )
@@ -128,7 +127,7 @@ def _parser() -> argparse.ArgumentParser:
     ingest.add_argument("--confirm-isolated-deployment")
 
     answer = commands.add_parser(
-        "answer", help="retrieve and call the frozen reader for one sample"
+        "answer", help="run the bounded public-recipe answer agent for one sample"
     )
     _run_and_sample(answer)
     answer.add_argument(
@@ -138,19 +137,20 @@ def _parser() -> argparse.ArgumentParser:
         help="run-absolute authorization; must cover the prepared tier item count",
     )
     answer.add_argument(
-        "--max-reader-calls",
+        "--max-agent-calls",
         type=int,
         required=True,
-        help="run-absolute ceiling over reader calls already recorded plus new calls",
+        help="run-absolute ceiling over answer-agent model calls; recipe calls"
+        " have a separate per-question cap",
     )
     answer.add_argument(
         "--max-evaluator-cost-usd",
         type=_positive_decimal,
         required=True,
-        help="run-absolute shared reader-plus-judge reported-cost ceiling",
+        help="run-absolute shared reported-spend stop threshold; use a provider"
+        " account cap as the hard monetary boundary",
     )
     answer.add_argument("--execute", action="store_true")
-    answer.add_argument("--confirm-index-ready")
 
     judge = commands.add_parser(
         "judge", help="call the frozen judge for one sample's answers"
@@ -166,7 +166,8 @@ def _parser() -> argparse.ArgumentParser:
         "--max-evaluator-cost-usd",
         type=_positive_decimal,
         required=True,
-        help="run-absolute shared reader-plus-judge reported-cost ceiling",
+        help="run-absolute shared reported-spend stop threshold; use a provider"
+        " account cap as the hard monetary boundary",
     )
     judge.add_argument("--execute", action="store_true")