fix: validate the pod name in emit_script, not only the CLI

[lesnik512]

Summary

Make the public emit_script validate its pod name, closing the deferred "harden the pod-cleanup trap" item. The pod name is embedded into shell in several places — most fragile the single-quoted EXIT trap (trap '... podman pod rm -f <pod> ... ; podman secret rm <pod>-<name> ...' EXIT) and the raw <pod>-<name> store names in secret create/secret rm. The CLI already validated it, but emit_script/EmitOptions are public API, so a library caller could pass an unvalidated pod and produce a malformed/unsafe trap.

Design: planning/changes/2026-07-12.01-validate-pod-name-in-emit.md.

Fix — validate the identifier, don't shell-escape it

A pod name is an identifier, not free text, so the honest-subset move is to reject a bad one (exactly how the codebase already treats secret/config names) rather than contort the trap to survive arbitrary metacharacters:

• POD_NAME_PATTERN (^[A-Za-z0-9][A-Za-z0-9_.-]*$) moves from cli.py to emit.py as the single source of truth; cli.py imports it.
• emit_script's first statement validates options.pod and raises UnsupportedComposeError on a non-match, guarding every caller (library and CLI).
• Both call sites use .fullmatch (not .match), closing the trailing-newline hole ("p\n" matches a $-anchored .match but breaks the trap) — the same lesson as the secret-name injection fix.
• No trap or escaping changes: a validated pod name is charset-inert, so shlex.quote is a no-op and the existing emission is provably safe.

Testing

just test-ci at 100% coverage (354 tests), just lint-ci clean, just check-planning OK. emit_script now raises on six adversarial pod values (single-quote injection, space, $(...), empty, leading dash, trailing newline); a valid pod still emits; the CLI's existing rejection (exit 2, same message) is unchanged.

An adversarial whole-branch review confirmed the security property is fully closed — the guard is emit_script's unconditional first statement, every char the pattern admits is in shlex.quote's safe class (so no nested quote ever enters the single-quoted trap), and both halves of every <pod>-<name> store name are validated.

Docs / planning

Retires the trap-hardening entry from deferred.md (its revisit trigger — a library-API test with adversarial pod values — is now met), and records the pod-name identifier constraint in architecture/supported-subset.md.