Add registry-based prod promotion, separate from auto dev deploy

Dev keeps auto-deploying on every push to main, same as before, but now
also publishes each build to git.jerodrigged.com's container registry
tagged by short commit SHA (best-effort - never blocks the dev deploy
if REGISTRY_TOKEN isn't set yet).

Prod deploys only on an intentional `git push origin main:prod`, and
only ever pulls a pre-built SHA-tagged image - it never rebuilds from
source. This guarantees prod runs the exact artifact dev already
validated, and makes promoting an untested commit fail loudly (pull of
a nonexistent tag) instead of silently rebuilding something new.

Needs new secrets before the prod path works: PROD_SSH_KEY, PROD_HOST,
PROD_USER, REGISTRY_USER, REGISTRY_TOKEN. PROD_SSH_KEY/PROD_HOST/
REGISTRY_USER are already set; PROD_USER and REGISTRY_TOKEN still need
Jerod's input.

CLAUDE.md

@@ -144,16 +144,43 @@ Missing images fall back to a styled placeholder — safe to deploy before photo
 | `.gitea/workflows/deploy.yml` | Push-to-main → SSH deploy → smoke tests |
 | `scripts/smoke-test.sh` | 6 curl assertions; exits non-zero on failure |
 
-## CI/CD — Trunk-Based Development
+## CI/CD — Dev auto-deploys, prod is a deliberate promotion
 
-Every push to `main`:
+Both pipelines run on Gitea Actions (git.jerodrigged.com). GitHub is not part
+of the deploy story — `main` on the GitHub mirror is unused.
+
+**Dev (every push to `main`)** — `.gitea/workflows/deploy.yml`:
 1. Gitea runner (CT117) SSHes into CT114 (dev LXC at 192.168.1.192)
 2. `git fetch origin && git reset --hard origin/main` — robust, never fails on drift
-3. `docker compose build && docker compose up -d`
-4. Health check loop (24 × 5s attempts)
-5. `bash scripts/smoke-test.sh` — 6 tests; pipeline fails if any fail
+3. `docker compose build` (tags images `lisilou-portfolio-web:local` / `lisilou-portfolio-api:local`)
+4. If `REGISTRY_TOKEN` is set: tag+push both images to
+   `git.jerodrigged.com/jhodgkin/lisilou-portfolio-{web,api}:<short-sha>` —
+   best-effort, a publish failure never blocks the dev deploy itself
+5. `docker compose up -d` (runs the just-built local image)
+6. Health check loop (24 × 5s attempts), then `bash scripts/smoke-test.sh`
 
-**Gitea secrets required:** `DEV_SSH_KEY`, `DEV_HOST`, `DEV_USER`
+**Prod (only on `git push origin main:prod`, or `<sha>:prod` to pin an older
+commit)** — `.gitea/workflows/deploy-prod.yml`:
+1. Gitea runner SSHes into CT111 (prod LXC, 192.168.1.246)
+2. `git fetch origin && git reset --hard origin/prod`, then computes the same
+   short SHA dev used to tag its published image
+3. `docker compose -f docker-compose.yml -f docker-compose.deploy.yml pull` —
+   pulls that exact SHA-tagged image from the registry. **No `docker compose
+   build` ever runs here** — this is what guarantees prod runs the same
+   artifact dev already validated, not a fresh rebuild that could drift.
+   Pulling a SHA that dev never published fails loudly instead of silently
+   rebuilding.
+4. `docker compose -f docker-compose.yml -f docker-compose.deploy.yml up -d`
+5. Same health check + smoke test pattern, against CT111
+
+`docker-compose.deploy.yml` is the override that swaps each service's `image:`
+for the registry-tagged one; it's only ever used for the prod pull, never for
+local dev (`docker compose up -d` alone still builds from source as before).
+
+**Gitea secrets required:** `DEV_SSH_KEY`, `DEV_HOST`, `DEV_USER` (dev);
+`PROD_SSH_KEY`, `PROD_HOST`, `PROD_USER` (prod); `REGISTRY_USER`,
+`REGISTRY_TOKEN` (both pipelines — a Gitea access token scoped
+`write:package,read:package`).
 
 ## Booking Wizard — Implementation Status