feat(compress): add provider model selection

[ousamabenyounes]

Fixes #759

Summary

• Add env-driven provider routing for caveman-compress: default Claude path, explicit anthropic, and explicit opencode.
• Add compress-specific model env support through CAVEMAN_COMPRESS_MODEL, with CAVEMAN_MODEL as fallback.
• Send opencode prompts through a temporary --file attachment instead of argv, then delete the temp file.
• Document provider/model env behavior and security implications.

CI will resync plugins/caveman/skills/caveman-compress/... on merge.

Test verification (RED -> GREEN)

RED, upstream behavior ignored opencode provider and still called Claude CLI:

FAIL: test_opencode_provider_uses_configured_model
Expected: run(['/usr/local/bin/opencode', 'run', '--model', 'github-copilot/gpt-4.1', 'Compress this memory.'], ...)
  Actual: run(['/usr/local/bin/opencode', '--print'], input='Compress this memory.', ...)
FAILED (failures=1, errors=1)

RED, Claude CLI ignored configured model:

FAIL: test_claude_cli_uses_configured_model
Expected: run(['/usr/local/bin/claude', '--model', 'claude-haiku-4-5', '--print'], ...)
  Actual: run(['/usr/local/bin/claude', '--print'], ...)
FAILED (failures=1)

RED, opencode path needed non-argv prompt attachment:

ERROR: test_opencode_provider_uses_configured_model
ValueError: '--file' is not in list
FAILED (errors=1)

GREEN, provider/model tests:

Ran 3 tests in 0.002s

OK

Full local CI baseline on upstream main:

# tests 112
# pass 112
python compress safety: Ran 5 tests ... OK

Full local CI final branch:

# tests 112
# pass 112
python compress safety: Ran 8 tests ... OK

Review gates:

contribute-code:     PASS
local-ci-docker:     PASS
pr-check:            PASS
ocr-codex:           PASS
main-codex-review:   PASS