feat(copilot): opus-context unique files (slimmed from 100-file bundle)

[arminanton]

Slimmed: copilot-opus-context unique files only (was a 100-file bundle)

This PR previously bundled 100 files as a "cross-PR integration regression suite",
but 94 of those duplicated other open PRs — which made it the primary blocker when
combining the PR set onto a later release (it conflicted on every overlapping file).

Slimmed to the 4 files genuinely unique to this PR:

hermes_cli/auth.py                                  # copilot-opus-context auth path
hermes_cli/runtime_provider.py                      # runtime provider resolution
tests/agent/conftest.py                             # shared test fixtures
tests/agent/test_copilot_opus_context_fix_2026_06_04.py   # the regression test

The 94 duplicate files are owned by their topical feature PRs already (autopilot
#49917, reasoning #48024, copilot identity #50064, etc.). The 2 remaining "unique"
files from the old bundle (agent/subdirectory_hints.py + its test) belong to the
RuntimeError-guard lineage and are covered by the superset #29433.

Built on v0.17.0 (2bd1977d8); all 4 files compile; 0 private-provenance leaks.
Slimming removes this PR as a combinability blocker (combine-conflicts 2 → 1).

[arminanton]

Closing: verification (build+test on a fresh v0.17.0 stack) proved this integration test cannot be green as a public PR. Its Phase B (agy-cli) and fable assertions depend on the agy-cli/fable rows in agent/models_dev.py _PROBE_VERIFIED_OVERRIDES, which are account-specific/private and deliberately NOT in any public PR (per project policy). On the public PR stack 25 of its tests fail (get_model_info('agy-cli',...) returns None); on the full tree (with the private data) all 64 pass. It therefore correctly belongs in the deferred private set (#50111), not a standalone public PR. Returning it there. This was caught by actually running the suite on the stacked tree rather than trusting git apply --check.

[arminanton]

Closing — content consolidated, intent covered, test not rebaseable

This PR is being closed (reversibly — the branch persists and it can be reopened) after exhaustive investigation. Summary of why, so nothing is lost:

This PR's base is v0.17.0; main is now 318 commits ahead. That staleness makes its diff misleading and its content non-viable as-is:

1. hermes_cli/auth.py (−670) and hermes_cli/runtime_provider.py (−194) are not genuine deletions — they are an artifact of diffing this PR's stale tree against current main. Landing them would revert main's auth improvements (env_int / _get_secret / profile-scoped helpers). The overlay's genuine change to these files vs v0.16.0 is small: the agy-cli ProviderConfig registration, which belongs with the isolated agy-cli PR (feat(provider): isolated agy-cli provider (WIP, draft — not for merge) #50555), not here.

2. tests/agent/conftest.py (the requires_ls_binary mark for agy-cli tests) has been relocated to feat(provider): isolated agy-cli provider (WIP, draft — not for merge) #50555 (commit on feat/agy-cli-provider-isolated), where the agy-cli tests that need it actually live. So that piece is preserved.

3. tests/agent/test_copilot_opus_context_fix_2026_06_04.py asserts the private overlay's exact copilot routing / effort-clamp / mythos-alias internals. It fails 55+ tests on every combination of current main + the relevant open PRs (fix(copilot): route Claude on Copilot to /v1/messages instead of /chat/completions #49184 + feat(reasoning): accept "max" reasoning effort end-to-end with nearest-down clamp #49644 + feat(provider): isolated agy-cli provider (WIP, draft — not for merge) #50555 were all applied and it still failed), because it tests overlay-specific values that the public PRs deliver differently. Its intent — verify copilot-claude → /v1/messages routing + effort clamps — is already delivered by:

   • main itself (copilot_model_api_mode routes claude → anthropic_messages)
   • fix(copilot): route Claude on Copilot to /v1/messages instead of /chat/completions #49184 (claude → v1messages routing, with its own passing tests)
   • feat(reasoning): accept "max" reasoning effort end-to-end with nearest-down clamp #49644 (max/effort clamp) and feat(models): correct under-reported Copilot/Codex per-model context and output limits #49449 (limits)

   Making it pass would require rewriting its assertions against current main (inventing new test code), which is out of scope for a faithful re-application PR.

Net: nothing the campaign needs on a later release is lost by closing — agy-cli code → #50555, the conftest → #50555, the test intent → main + #49184/#49644/#49449. The agy-cli auth/runtime registration can be added to #50555 (the agy-cli holding pen) at upgrade time; it applies cleanly from v0.16.0.

Reopen if a fresh, current-main-based opus-context test is ever wanted.