20260816 - Resolve the telemetry image tag when building the artifact

[Purple10101]

Adding retina-telemetry to docker-compose.yml was not enough to ship it.

The bug

The artifact build resolves image tags with one hand-written sed clause per variable, and TELEMETRY_V had none. The generated manifest kept a literal:

image: ghcr.io/offworldlabs/retina-telemetry:${TELEMETRY_V:-v0.1.0}

Which is exactly what the comment above that block warns about:

gen_docker-compose uses raw sed to extract image names and cannot handle unresolved shell variables

skopeo receives ${TELEMETRY_V:-v0.1.0} as a tag. The release either fails there, or produces an artifact whose telemetry service can never start on a node.

Why nothing caught it

The compose file is valid. docker compose config resolves the variable correctly. The same service block ran on two live nodes against production. The defect isn't in the compose file at all — it's in a transformation of it that enumerates its variables by hand, in another file, with no test.

The fix

The seventh clause, plus a guard that removes the need to remember an eighth: every image line in the generated manifest must be a literal, and the build fails naming the variable that was missed.

Verified against the merged docker-compose.yml:

	before	after
image tags resolved	9 of 10	10 of 10
guard on the old six-clause build	—	exits 1, printing line 178 and the variable
runtime ${} refs preserved	—	35, untouched

Runtime variables are deliberately left alone and still resolved from .env at deploy time — the check reads image: lines only. bash -n clean; generated manifest parses as YAML with all 10 services.

🤖 Generated with Claude Code