Shared hivemind provider's pg_hba refuses Polish fleet hosts (gateway source 100.64.0.10), so tavern cannot run there

Problem and intended result

The shared hivemind provider (.aidev/shared-hivemind.compose.yml, served at 192.168.3.193:44149) refuses every tavern run from the Polish fleet hosts:

FATAL: no pg_hba.conf entry for host "100.64.0.10", user "hivemind_shared_consumer", database "hivemind_shared", no encryption

Seen by the #416 (closed) worker and again by the #418 (closed) worker (2026-10-25). Polish hosts reach the provider through a gateway that rewrites their source address to 100.64.0.10. PG_ACCESS (.aidev/shared-hivemind.compose.yml, the PG_ACCESS: line) admits 192.168.3/5/6.0/24 and 172.16.0.0/12 only, so the tavern step of the full gate fails or cannot run whenever a hivemind workflow or canary lands on a Polish host. #418 (closed) showed tavern passing (19 groups) from steem-5, a US host, on the same tree.

Intended result: tavern reaches the shared provider from Polish hosts exactly as it does from US hosts.

Required behavior

  1. Add 100.64.0.10/32 to PG_ACCESS with the same two rules each existing subnet has: hivemind_shared for hivemind_shared_consumer, and the /^run_[a-z0-9_-]+$$ databases for hivemind_shared_consumer,hivemind. Use the single gateway address (/32), not the whole 100.64.0.0/10 range: these rules are trust, and nothing else on that range should get it.
  2. Update the comment above PG_ACCESS to say why the address is there (the Polish-fleet gateway's source rewrite).
  3. The running provider only reads PG_ACCESS when its container is created. Find how this provider's hosting session applies a compose change (the comments at the top of the file and .aidev/README.md) and either do that or state on this issue the exact operator step needed, with the command.

Acceptance criteria

  • Under the updated PG_ACCESS, a connection to hivemind_shared as hivemind_shared_consumer from source 100.64.0.10 is accepted, and from an address outside every listed range is still refused.
  • Under the updated provider, the full gate's tavern step run from a Polish fleet host connects to the provider and its groups run (any test failures are reported as test failures, not a pg_hba refusal).
  • The existing subnets' rules are unchanged.

Verification

Render the resulting pg_hba entries from PG_ACCESS and check them. If the provider can be recreated from this workflow, run tavern from a Polish host against it. If not, record the operator step on this issue: the live check is the next hivemind gate that lands on a Polish host, and the lead will watch for it.

Integration points

  • .aidev/shared-hivemind.compose.yml — PG_ACCESS gains the two 100.64.0.10/32 rules and the comment explaining them

Scope boundary

The gateway's address rewriting is network configuration outside this repository and is not changed.

Related-To: #418 (closed)