fix(env): fail-closed Cloud Agent start and hash-locked install

.cursor/environment.json

@@ -0,0 +1,8 @@
+{
+  "name": "Codec Carver",
+  "install": "bash .cursor/install.sh",
+  "start": "bash .cursor/start.sh",
+  "ports": [
+    { "name": "web", "port": 8000 }
+  ]
+}

.cursor/install.sh

@@ -0,0 +1,41 @@
+#!/usr/bin/env bash
+# Idempotent Cloud Agent setup for codec-carver.
+# Runs after each checkout. Always resolve the repository root from this file
+# so a cached or non-root working directory cannot pip/cargo the wrong tree.
+set -euo pipefail
+cd "$(dirname "$0")/.."
+export PATH="${HOME}/.local/bin:${PATH}"
+
+# ffmpeg/ffprobe are required at runtime for probing and conversion.
+# The default Cloud Agent base image already ships them; install only if missing.
+if ! command -v ffmpeg >/dev/null 2>&1; then
+  sudo apt-get update
+  sudo apt-get install -y --no-install-recommends ffmpeg
+fi
+
+# rustfmt backs `cargo fmt` (rust-toolchain.toml pins a minimal profile without it).
+# Fail closed: CI and local agents both run `cargo fmt --check`.
+if ! command -v rustup >/dev/null 2>&1; then
+  echo "rustup is required so rustfmt can be added for cargo fmt checks." >&2
+  exit 1
+fi
+rustup component add rustfmt
+
+# Match CI exactly: install the authenticated hash lock, then link the checked-out
+# editable tree without dependency resolution or index access. requirements-dev.txt
+# remains an opt-in local fuzzing surface until it has its own reviewed hash lock;
+# unattended Cloud Agent bootstrap must not resolve those unhashed requirements.
+python3 -m pip install --user --disable-pip-version-check \
+  --require-hashes -r requirements-lock.txt
+python3 -m pip install --user --disable-pip-version-check \
+  --no-index --no-deps --no-build-isolation -e .
+
+# Rust core binary (codec-carver-core) used by the audio-library CLI.
+# A fresh checkout wipes rust-core/target, so (re)build it here.
+cargo build --release --manifest-path rust-core/Cargo.toml
+
+# Expose pip --user console scripts (codec-carver, codec-carver-library, ...) in
+# non-login shells too. Login shells already pick up ~/.local/bin via ~/.profile.
+if ! grep -q '.local/bin' "$HOME/.bashrc" 2>/dev/null; then
+  printf '\nif [ -d "$HOME/.local/bin" ]; then PATH="$HOME/.local/bin:$PATH"; fi\n' >> "$HOME/.bashrc"
+fi

.cursor/start.sh

@@ -0,0 +1,48 @@
+#!/usr/bin/env bash
+# Per-boot reconciliation. install is not re-run when a pod boots from a
+# prebuilt environment, so anything that must exist on every boot lives here.
+# `start` must reach a clear ready-or-fail state: a green exit with a dead
+# uvicorn leaves the published :8000 port pointing at nothing.
+set -euo pipefail
+cd "$(dirname "$0")/.."
+export PATH="${HOME}/.local/bin:${PATH}"
+
+WEB_HEALTH_URL="${CODEC_CARVER_WEB_HEALTH_URL:-http://127.0.0.1:8000/health}"
+WEB_READY_ATTEMPTS="${CODEC_CARVER_WEB_READY_ATTEMPTS:-40}"
+WEB_LOG="${CODEC_CARVER_WEB_LOG:-/tmp/codec-carver-web.log}"
+
+# 1) Ensure the Rust backend binary exists (a fresh checkout wipes rust-core/target).
+if [ ! -x rust-core/target/release/codec-carver-core ]; then
+  cargo build --release --manifest-path rust-core/Cargo.toml
+fi
+
+_web_is_ready() {
+  curl -sf -o /dev/null "${WEB_HEALTH_URL}"
+}
+
+# 2) Start the FastAPI SaaS web service (saas_web:app) if it is not already up.
+# Backgrounded so `start` returns after readiness; the probe keeps it a single
+# instance on every boot mode (just-in-time or prebuilt build/snapshot).
+if _web_is_ready; then
+  exit 0
+fi
+
+nohup python3 -m uvicorn saas_web:app --host 0.0.0.0 --port 8000 \
+  > "${WEB_LOG}" 2>&1 &
+web_pid=$!
+
+attempt=0
+while [ "${attempt}" -lt "${WEB_READY_ATTEMPTS}" ]; do
+  if _web_is_ready; then
+    exit 0
+  fi
+  if ! kill -0 "${web_pid}" 2>/dev/null; then
+    echo "uvicorn exited before becoming ready; see ${WEB_LOG}" >&2
+    exit 1
+  fi
+  attempt=$((attempt + 1))
+  sleep 0.25
+done
+
+echo "uvicorn did not become ready after ${WEB_READY_ATTEMPTS} probes; see ${WEB_LOG}" >&2
+exit 1

AGENTS.md

@@ -85,6 +85,17 @@ repo.
   or feature-extraction path.
 <!-- END cwl-agent-guidance -->
 
+## Cloud Agent environment
+
+`.cursor/environment.json` is repository-managed and overrides dashboard
+environments. `install` (`bash .cursor/install.sh`) must stay idempotent,
+hash-lock Python deps, and fail if `rustfmt` cannot be added. `start`
+(`bash .cursor/start.sh`) must wait on `GET /health` before exiting 0; do not
+`nohup` uvicorn and return. Do not add `$schema` to the JSON. Decision record:
+[`docs/doctoring/cloud-agent-environment.md`](docs/doctoring/cloud-agent-environment.md).
+The module map, job-store ERD, and connector order live in
+[`ARCHITECTURE.md`](ARCHITECTURE.md).
+
 ## Code-owner review gates — disabled (on hold)
 
 As of 2026-08-04, code-owner review requirements (`require_code_owner_reviews` in branch

ARCHITECTURE.md

@@ -0,0 +1,134 @@
+# Architecture
+
+Codec Carver is the speech/video conversion module for STT and omni-modal LLM
+input. It must run **standalone** (CLI, FastAPI upload service, MCP tool) and
+as a **git submodule** consumed by naruon and sibling ContextualWisdomLab
+services.
+
+## Runtime surfaces
+
+```mermaid
+flowchart LR
+    subgraph standalone [Standalone]
+        CLI["codec-carver CLI"]
+        WEB["saas_web FastAPI"]
+        MCP["mcp_driver shrink_media"]
+    end
+    subgraph core [Conversion core]
+        MS["media_shrinker.convert_file"]
+        RS["rust-core codec-carver-core"]
+    end
+    CLI --> MS
+    WEB --> MS
+    MCP --> MS
+    CLI --> RS
+    WEB --> JOB["job_store SQLite WAL"]
+    MS --> FF["ffmpeg / ffprobe"]
+```
+
+| Surface | Module | Buyer action |
+| --- | --- | --- |
+| CLI | `media_shrinker.py` | Carve a folder of recordings into size-capped FLAC/Opus. |
+| Library CLI | `audio_library.py` → Rust | Inventory, hash, TMK, and GPU transcription batches. |
+| Web | `saas_web.py` | Upload one file or a batch; poll `/jobs/{id}` for long work. |
+| MCP | `mcp_driver.py` | Call `shrink_media` from an agent runtime. |
+| Jobs | `job_store.py` | Survive process restart; do not keep results only in RAM. |
+
+Sources are never overwritten. Generated names keep the original filename plus
+a new suffix (`meeting.wav.flac`).
+
+## Conversion pipeline
+
+```mermaid
+flowchart TD
+    A[find_candidates] --> B[probe_media / ffprobe]
+    B --> C{duration over cap?}
+    C -->|yes| D[silencedetect + build_segments]
+    C -->|no| E[single segment]
+    D --> F[build_audio_plan FLAC]
+    E --> F
+    F --> G{_execute_plan}
+    G -->|FLAC over target| H[build_opus_plan]
+    H --> G
+    G --> I[preserve_file_attributes]
+    I --> J[write_report JSON]
+```
+
+Silence splits prefer long quiet intervals; a hard split just under the
+duration cap is the fallback. Parsers of ffmpeg/ffprobe text
+(`parse_silencedetect_intervals`, `_parse_probe_payload`) treat that text as
+untrusted input and raise `MediaShrinkerError` rather than unexpected types.
+
+## Cloud Agent environment
+
+`.cursor/environment.json` is the highest-precedence environment source for
+Cloud Agents. Do not add `$schema` (the public schema rejects undeclared
+fields).
+
+| Phase | Script | Must do |
+| --- | --- | --- |
+| `install` | `.cursor/install.sh` | `cd` to repo root; hash-locked pip; fail-closed rustfmt; release-build Rust. |
+| `start` | `.cursor/start.sh` | Rebuild Rust only if the binary is missing; start uvicorn; **wait on `GET /health`**. |
+| ports | `8000` / `web` | Bind `0.0.0.0:8000` so the published port reaches the SaaS UI. |
+
+Decision record: [`docs/doctoring/cloud-agent-environment.md`](docs/doctoring/cloud-agent-environment.md).
+
+## Job-store data (current)
+
+`job_store.py` persists async conversions. Table `jobs` is a single-word name
+and is **known debt**: the org naming rule requires two-or-more-word
+snake_case objects (`conversion_jobs`) plus a 3NF-preserving rename migration.
+Do not add more single-word tables. Columns already use multi-word names where
+they store paths (`output_path`, `output_name`, `temp_dir`).
+
+```mermaid
+erDiagram
+    JOBS ||--o| CONVERSION_OUTPUT : produces
+    JOBS {
+        text id PK
+        text status
+        text created_at
+        text updated_at
+        text output_path
+        text output_name
+        text error
+        text temp_dir
+    }
+```
+
+Callers pass `now` explicitly. The store never calls `datetime.now()`.
+
+## Security and operability baseline
+
+- ffmpeg/ffprobe: `-nostdin` and `-protocol_whitelist file,crypto,data`.
+- Uploaded names: sanitize to a safe basename; normalize Windows separators.
+- Permissions copied onto outputs drop setuid/setgid/sticky bits.
+- Opt-in API keys (`CODEC_CARVER_API_KEYS`) are a **known deviation**. Runtime
+  reads must move to a SQLite/KV credential registry; env is bootstrap
+  transport only (issues #329, #373).
+- `GET /` and `GET /health` stay reachable without a key so a buyer can open
+  the form and a probe can confirm liveness.
+- PII in recordings is the product. Do not mask audio content; protect it with
+  access control, retention, and encryption at rest instead.
+
+## Ecosystem connectors (leverage order)
+
+1. **naruon** — consume carved FLAC/Opus + JSON reports as speech/DOM input.
+2. **contextual-orchestrator** — route transcription/LLM describe calls; do not
+   read provider keys from raw env at runtime.
+3. **wardnet** — place the SaaS UI behind the org WAF/APIM when exposed.
+4. **keyverse** — replace shared API keys with passwordless service identity.
+5. **clearfolio / newsdom-api** — attach transcripts and reports to documents.
+
+Each connector must keep this repo independently runnable.
+
+## Test and research baseline
+
+- Unit/integration: `python3 -m unittest discover -s tests -v` (100% coverage
+  over the configured sources; 100% interrogate on production modules).
+- Reality checks: WAV → FLAC lossless; tiny `--target-bytes` → Opus; silence
+  splits on long fixtures; Rust inventory SHA-256.
+- Fuzz: Atheris harnesses in `fuzz/` plus Hypothesis mirrors in
+  `tests/test_fuzz_properties.py`.
+- Speech/codec changes attach APA 7th citations under `docs/doctoring/` and
+  `docs/papers/` (PDF only when redistribution is permitted).

CHANGELOG.md

@@ -2,6 +2,8 @@
 
 ## [Unreleased]
 ### Added
+- Cloud Agent `install`/`start` 스크립트가 해시 잠금 pip, fail-closed rustfmt, `GET /health` 준비 대기를 강제합니다. 근거는 [`docs/doctoring/cloud-agent-environment.md`](docs/doctoring/cloud-agent-environment.md)에 있습니다.
+- `GET /health` 활성 프로브를 추가해 업로드 페이지와 같이 API 키 없이 프로세스 생존을 확인할 수 있습니다.
 - 다중 파일 업로드 선택 시 즉각적인 파일 개수 피드백 및 제한 초과 경고 메시지 추가
 - 일괄 업로드 폼에 대상 바이트 프리셋 버튼과 총 파일 크기 미리보기를 추가하여 사용성을 개선했습니다.
 - 클라이언트 측 폼 검증 시 하드코딩된 '5 GiB' 텍스트를 동적으로 변환되도록 수정하고 일괄 업로드 폼에 최대 크기(MAX_UPLOAD_BYTES) 검증 피드백을 추가했습니다.

CLAUDE.md

@@ -47,9 +47,10 @@ CI installs pinned dependencies with `pip install --require-hashes -r requiremen
 Four flat top-level modules (declared as `py-modules` in `pyproject.toml`; there is no package directory):
 
 - **`media_shrinker.py`** — the core engine and CLI, deliberately stdlib-only (external work happens in `ffmpeg`/`ffprobe` subprocesses). The console script `codec-carver` maps to `media_shrinker:main`. Pipeline for a batch run: `find_candidates` scans the root (pruned `os.walk`, excludes the output dir and `--exclude-dir-prefix` dirs) → per file, `convert_file` probes with ffprobe (`probe_media` / `_parse_probe_payload`), detects silence and builds a split plan for long sources (`detect_silence_intervals`, `parse_silencedetect_intervals`, `build_segments`) → each segment gets a `ConversionPlan` (`build_audio_plan` prefers FLAC; `build_opus_plan` is the fallback when a FLAC output exceeds the target size) → `_execute_plan` runs ffmpeg and `preserve_file_attributes` restores permissions/timestamps/xattrs best-effort → `write_report` emits a JSON report. `convert_file(source, root=..., output_dir=..., target_bytes=...)` is the programmatic API that the web and MCP layers call.
-- **`saas_web.py`** — single-file FastAPI upload UI (the `[web]` extra; what the Docker image serves). Streams one upload into a temp workspace, calls `media_shrinker.convert_file`, and returns the first generated output as a download. Middleware enforces a 5 GiB upload cap and security headers. Processing is synchronous per request.
+- **`saas_web.py`** — single-file FastAPI upload UI (the `[web]` extra; what the Docker image serves). Streams one upload into a temp workspace, calls `media_shrinker.convert_file`, and returns the first generated output as a download. `GET /health` is the liveness probe used by Cloud Agent `start` and load balancers. Long work uses `/jobs` backed by `job_store`. Middleware enforces a 5 GiB upload cap, optional API keys, and security headers.
 - **`mcp_driver.py`** — FastMCP server (the `[mcp]` extra) exposing a single `shrink_media` tool that wraps `convert_file`.
-- **`job_store.py`** — stdlib-only SQLite (WAL) durable job store intended for async/worker job tracking. It is tested but not yet wired into `saas_web.py`. Callers pass `now` explicitly; the store never calls `datetime.now()` itself.
+- **`job_store.py`** — stdlib-only SQLite (WAL) durable job store used by `saas_web.py` async jobs. Callers pass `now` explicitly; the store never calls `datetime.now()` itself.
+- **`.cursor/`** — repository-managed Cloud Agent environment. `install.sh` hash-locks pip and fail-closes rustfmt; `start.sh` waits on `GET /health`. See `ARCHITECTURE.md` and `docs/doctoring/cloud-agent-environment.md`.
 
 Supporting directories: `fuzz/` holds Atheris harnesses plus seed corpora for the three untrusted-input parsing surfaces (`parse_silencedetect_intervals`, `_parse_probe_payload`, `build_segments`); the same invariants run as Hypothesis property tests in `tests/test_fuzz_properties.py` so they execute in the normal suite. `docs/papers/` holds the fuzzing survey the harness design references.
 

Dockerfile

@@ -22,8 +22,8 @@ USER appuser
 
 EXPOSE 8000
 
-# Liveness: the FastAPI app serves the upload UI at "/".
+# Liveness: use the same cheap JSON readiness surface as Cloud Agent startup.
 HEALTHCHECK --interval=30s --timeout=5s --start-period=10s --retries=3 \
-    CMD python -c "import urllib.request,sys; sys.exit(0 if urllib.request.urlopen('http://127.0.0.1:8000/', timeout=4).status==200 else 1)"
+    CMD python -c "import urllib.request,sys; sys.exit(0 if urllib.request.urlopen('http://127.0.0.1:8000/health', timeout=4).status==200 else 1)"
 
 CMD ["uvicorn", "saas_web:app", "--host", "0.0.0.0", "--port", "8000"]

docs/doctoring/cloud-agent-environment.md

@@ -0,0 +1,68 @@
+# Cloud Agent environment bootstrap
+
+## Decision
+
+Codec Carver commits a repository-managed Cloud Agent environment at
+`.cursor/environment.json`. `install` installs hash-locked Python dependencies,
+adds `rustfmt` to the pinned minimal Rust toolchain, and builds
+`codec-carver-core`. `start` rebuilds that binary only when a fresh checkout
+wiped `rust-core/target`, then brings up `saas_web:app` on `0.0.0.0:8000` and
+**waits** until `GET /health` succeeds.
+
+Agents should treat a green `start` as “the upload UI is listening.” A
+`nohup` without a readiness wait is not an environment start.
+
+## Technical basis
+
+Cursor resolves environment configuration from the checked-out
+`.cursor/environment.json` before any dashboard-managed personal or team
+environment (Cursor, n.d.). The public schema rejects undeclared fields,
+including `$schema`, so this repository does not add one.
+
+Python package integrity follows pip hash-checking mode: install the lock
+file with `--require-hashes`, then link the editable tree with `--no-index
+--no-deps` so the index cannot substitute a different wheel (Python Software
+Foundation, n.d.). That pairing is the same contract CI uses and is the
+minimum SLSA-aligned control this repository can apply without a hermetic
+build service (SLSA, 2023).
+
+Fail-closed toolchain setup follows NIST SSDF PW.4 / PW.8: produce a
+repeatable build environment and do not continue when a required verification
+tool (`rustfmt`) is missing (Souppaya et al., 2022).
+
+`GET /health` is a liveness probe, not a substitute for job-store or ffmpeg
+readiness. Load balancers and Cloud Agent `start` share that URL so a crashed
+uvicorn cannot look like a successful boot.
+
+## Verification and rollback
+
+- `bash -n .cursor/install.sh .cursor/start.sh` must succeed.
+- `tests/test_cloud_agent_environment.py` locks the JSON fields, the hash-locked
+  pip lines, rustfmt fail-closed install, and the `/health` wait.
+- `GET /health` stays auth-exempt and returns
+  `{"status": "ok", "service": "codec-carver"}` whether or not API keys are
+  configured.
+- Roll back by restoring the previous scripts; do not reintroduce `|| true` on
+  rustfmt or a `nohup` that returns before the probe succeeds.
+
+## Next action
+
+After this lands, mark the Cloud Agent environment draft ready for review and
+close the earlier environment-only PR that lacked the readiness contract.
+
+## References
+
+Cursor. (n.d.). *Cloud Agents environment schema*. Retrieved August 16, 2026,
+from https://cursor.com/schemas/environment.schema.json
+
+Python Software Foundation. (n.d.). *Secure installs: Hash-checking mode*
+(pip documentation). Retrieved August 16, 2026, from
+https://pip.pypa.io/en/stable/topics/secure-installs/
+
+SLSA. (2023). *Supply-chain Levels for Software Artifacts (SLSA) v1.0*.
+https://slsa.dev/spec/v1.0/
+
+Souppaya, M., Scarfone, K., & Dodson, D. (2022). *Secure Software Development
+Framework (SSDF) Version 1.1: Recommendations for mitigating the risk of
+software vulnerabilities* (NIST Special Publication 800-218). National
+Institute of Standards and Technology. https://doi.org/10.6028/NIST.SP.800-218