fix(ci): repair v0.18.2 merge corruption (dropped fork symbols/columns/tests) → green main

[OmarB97]

Problem

Fork main is red across all CI (ruff invalid-syntax blocks lint; test collection fails). Root cause: the merge commit ba510aab2e ("Merge upstream Hermes v0.18.2 into fork") resolved conflicts by systematically favoring the upstream side and dropping fork-side additions while keeping references to them — leaving dangling calls, dropped columns, and stitch-corrupted test files.

The two broken test files couldn't even be collected, so the failures had been latent (never run) since the merge.

What changed — every corruption found and repaired

Repaired from known-good sources: fork parent e5d67be7 (all fork additions, parses clean), upstream v0.18.2 (31e39dec84), and the deployed/working hermes_state.py (same SCHEMA_VERSION 19).

hermes_state.py

• create_session INSERT dropped the device_name column + placeholder + ON CONFLICT clause (13-vs-14 binding crash). Restored (now matches deployed).
• messages INSERT dropped sender_device (17-vs-18 binding crash). Restored.
• SessionDB._surfaced_session_clause (staticmethod, 3 call sites in list_sessions_rich, dep _delegate_from_json present) — definition dropped. Restored.
• SessionDB._batch_compression_lineages (call site in the compression-roots branch) + _batch_compression_tips (direct test coverage) — dropped. Restored.
• list_sessions_rich non-compression-roots branch lost its WITH msg_max AS (...) CTE (kept the newer _compact_session_cols _sel, only the CTE definition was dropped → no such table: msg_max). Restored.

hermes_cli/models.py

• model_switch.py's list_authenticated_providers imports/calls cached_api_model_metadata (the fork's live-router-metadata feature), but the merge dropped BOTH the cache-population block in fetch_api_models AND the cached_api_model_metadata accessor — a dangling from hermes_cli.models import cached_api_model_metadata that raised ImportError on every custom-provider probe path. Restored both, preserving upstream's fetch_api_models(request_headers=headers) signature.

Test files (tests/test_hermes_state.py, tests/hermes_cli/test_model_switch_custom_providers.py)

• Both had a merge stitch that spliced a fork test into an adjacent upstream test, producing a syntax error. Reconstructed: fork mid99 test regained its assertions and upstream TestListCronJobRuns its class header; the fork carry_live_router_metadata probe test and upstream discover_models_false test were separated back into two functions.

How to review

1. hermes_state.py — the two INSERTs (device_name/sender_device), the two grafted compression helpers + _surfaced_session_clause before list_sessions_rich, and the restored WITH msg_max CTE.
2. hermes_cli/models.py — fetch_api_models cache population + cached_api_model_metadata accessor.
3. The two test files — the reconstructed function boundaries.

Evidence

• Syntax sweep: every .py in the repo compiles (py_compile), so ruff invalid-syntax is cleared.
• python -m pytest tests/test_hermes_state.py -q → 346 passed (was uncollectable → 304 → 344 → 346 as each dropped symbol was restored).
• from hermes_cli.models import cached_api_model_metadata now imports (was ImportError); list_authenticated_providers imports clean.
• Comprehensive dropped-symbol scan (fork parent vs main, dangling references only) surfaced these and no others once refactor false-positives (e.g. _add_column_if_missing → moved to hermes_cli.sqlite_util; _git_branch_for_cwd = git_probe.branch) were excluded.

Verification

• python -m pytest tests/test_hermes_state.py -q → 346 passed. PASS
• python -m py_compile over the whole tree → clean. PASS
• Remaining 4 test_model_switch_custom_providers.py failures in the local sandbox are the model-catalog fetch (hermes-agent.nousresearch.com/docs/api/model-catalog.json → HTTP 403 in a network-restricted sandbox) changing model counts — environmental. The two non-reconstructed ones (total_models_reflects_grouped_count, distinct_endpoints_stay_separate) are upstream tests that passed in the paired upstream PR's CI; CI here (with network) is the arbiter for the router-metadata probe tests.

Risks / gaps

• One further upstream-side param, force_fresh_nous_tier in model_switch.py, was also dropped by the merge (fork-main hardcodes _nous_free(force_fresh=True)). It does not crash — it just always forces a fresh Nous-tier check. Left as-is unless CI flags a test for it; noted on task hermes-fork-v0182-merge-botched-redo. Not a blocker for green CI.
• The comprehensive fix here is scoped to dangling/broken fork drops (what breaks CI). Any silently-dropped fork feature with no test/reference is out of scope; the merge should still be reviewed holistically (tracked on the same task).