fix(cold-storage): create cold directory before archiving idle models

[vaibhavsrv]

Why this matters

In ods/scripts/llm-cold-storage.sh, when cold storage archiving is executed on a host where $COLD_DIR does not yet exist, mv "$model_dir" "$COLD_DIR/$name" fails because the target directory has not been created. Because the move failure was unhandled, the script proceeded to execute ln -s "$COLD_DIR/$name" "${model_dir%/}" against the unmoved source directory (creating a broken nested link inside the intact cache directory) and unconditionally incremented the archived counter, falsely reporting successful archive operations while leaving models in place.

This patch ensures $COLD_DIR is created before attempting to move idle models, guards the subsequent symlink creation and success counters on mv exiting 0, and ensures $HF_CACHE exists prior to restoring archived models. Existing model retention thresholds, protected model sets, and symlink resolution contracts remain unchanged.

Validation

• Repro baseline: Invoked llm-cold-storage.sh --execute against an idle model directory with a non-existent $COLD_DIR; confirmed mv failed, the model was left unarchived, and an erroneous symlink was created inside the cache directory while reporting 1 archived.
• Post-fix: Verified test_cold_storage_initial_directory.py passes with exit code 0; $COLD_DIR is created, the model is cleanly moved, symlink resolution verifies against the archived payload, and restore roundtrips the directory back to $HF_CACHE.
• Cold storage suites: 1 passed. New-test Ruff, ShellCheck, and diff checks pass; new regressions wired into Linux CI.

Overlap check

• fix(env): probe .env keys literally in _env_set and mode-switch env_set #5819 (0xacee): Addressed macOS environment generation and safe escaping; touches ods/installers/macos/ and has zero overlap with llm-cold-storage.sh.
• fix(scripts): validate context length, port bounds, and control characters in render-runtime-configs #5056 (upstream): Touched model store contracts in Compose; does not alter cold storage archive scripts or cache migration paths.
• fix(dashboard-api): tolerate evicted unresolved attempts during pixel chat result retrieval #5699 (vaibhavsrv): Addressed token provider boundary tests without touching storage archiving.

Risk / AI disclosure

AI-assisted investigation, implementation and CLI regressions. This strengthens filesystem directory creation and return-code guards during model storage tiering, not runtime model execution. Independent human review and platform/runtime qualification remain gates. No running configuration, deployment or upstream merge changed.

Follow-up integration evidence

Composed with #5699 at 65a2a7a51 without conflicts. Production and test diffs passed together; cold storage relative-path round trips and preservation checks remain intact.
Backlog composition was local-only (production/test diffs, excluding workflow/Makefile wiring); it is not an upstream merge or independent human approval. Declared live-review gates remain open.

[Lightheartdevs]

Thanks for this contribution. public-beta was promoted into main on 2026-09-24 and no longer receives changes, so we're closing pull requests that target it. This isn't a judgment on the change itself. If it's still needed, please rebase onto main and open a focused PR. See #7253 for details and the contribution policy.