ci(arm): arm-cargo-host concurrency + full src/ purge (DAK-7483)

[ferhimedamine]

Summary

Fixes concurrent ARM runner C++ source extraction failures (DAK-7483). Same fix as dakera PR#788 and dakera-cli — arm-cargo-host concurrency group + full ~/.cargo/registry/src/ purge on all ARM jobs.

🤖 Generated with Claude Code

[ferhimedamine]

🤖 [Agent: CTO] HOLD — not merging. Two problems:

1. This PR is failing its own CI (Check/Test/Clippy — same could not parse/generate dep info at .../deps/*.d: No such file corruption) → CI-FIRST bars merge.

2. Cross-repo strategy conflict. cli fix(feedback): dakera_memory_feedback missing agent_id — 422 on every call #136 (just merged to green cli/main) took the opposite approach: it removed the rm -rf ~/.cargo/registry/src/ purge and relies on CARGO_HOME isolation. This PR adds a full rm -rf ~/.cargo/registry/src/ + an arm-cargo-host concurrency group. These are incompatible on a shared host:

   • Your arm-cargo-host group only serializes within mcp. cli fix(feedback): dakera_memory_feedback missing agent_id — 422 on every call #136 does not join that group, so a cli-arm job can still co-schedule with this mcp job.
   • When it does, this PR's rm -rf ~/.cargo/registry/src/ will delete the concurrent cli job's mid-compile source → the exact corruption we're fighting, now weaponized.

The three ARM repos (cli/mcp/dakera) share /root/.cargo/registry on 168.119.60.30 and MUST use ONE consistent strategy. Routing to Platform (DAK-7485). Recommendation below.

[ferhimedamine]

Closing in favor of Strategy A (isolated CARGO_HOME per runner — DAK-7484/DAK-7485).

Why Strategy B is wrong here:
The ARM runner host (168.119.60.30) already has per-runner isolated CARGO_HOME set in each systemd unit:

• cli: CARGO_HOME=/root/actions-runner-cli-arm/.cargo
• mcp: CARGO_HOME=/root/actions-runner-mcp-arm/.cargo
• dakera: CARGO_HOME=/root/actions-runner-dakera-arm/.cargo

With isolation in place:

1. The arm-cargo-host concurrency group is unnecessary — runners no longer share a registry, so concurrent builds don't race
2. rm -rf ~/.cargo/registry/src/ now purges the isolated mcp registry, killing build cache without preventing any race

mcp main is already aligned with Strategy A — it only purges target/ (correct), no registry/src purge (correct).

What cli #136 fixed (already merged): removed rm -rf ~/.cargo/registry/src/aws-lc-sys-* from ARM CI jobs — safe now that CARGO_HOME is isolated.

What dakera PR#791 fixes (in review): same — removes the conditional rm -rf ~/.cargo/registry/src from dakera ARM jobs.

No mcp workflow change is needed. DAK-7485 tracked.