chore/docs: cleanup committed root artifacts + minor docs/diagnostics fixes

[EAbaracus]

Summary

Small, low-risk hygiene PR (from a static read-only audit of the repo) covering 5 "good first" items:

• Remove committed root artifacts (HA-D11-001 / HA-D3-001): log.txt (empty 0-byte), sqlite_leak_fix.png (832 KB, unreferenced), default.tar.gz (1.96 MB — only ever written by tests/hermes_cli/test_profiles.py into a temp dir, never read from repo root). Added ignore rules so they cannot be re-committed.
• Diagnostics (HA-D6-001): acp_adapter/server.py fell back to HERMES_VERSION = "0.0.0" silently on import failure; now logs a warning via the existing module logger.
• Docs (HA-D12-001 / HA-D9-001 / HA-D2-001):
  • AGENTS.md: convert the ui-tui dev-command inline bash snippet to a numbered, copy-paste-friendly list.
  • README.md: add an AGENTS.md cross-link on the Windows install note so contributor/dev setup has a single source of truth.
  • agent/secret_sources/base.py: document the deliberate shell=True exception in command_token_source (operator's own key_cmd), resolving the apparent contradiction with the no-shell rule.

Test plan

• git rebase upstream/main — clean, no conflicts
• git diff --stat upstream/main.. shows only the intended files
• Confirmed the 3 root artifacts still exist on upstream/main (deletion is legit cleanup, not accidental)
• CI (tests.yml / lint.yml) should pass — changes are docs + 1 log line + file deletion

Notes

Not security-critical; no behavioral change except the new warning log. Branch is on the EAbaracus fork; base is NousResearch/hermes-agent@main.

[Enough1122]

AI code review — automated review for reference; please use your judgment.

1. .gitignore:100–105 — the new patterns are unanchored (log.txt, sqlite_leak_fix.png, default.tar.gz), so they ignore files with those names at every depth, not just repo root — why it matters: log.txt in particular is a plausible name for a test fixture or plugin sample data; when someone adds one later, git will silently refuse to track it and the failure mode ("why won't my file commit?") wastes an hour — suggestion: anchor them (/log.txt, /sqlite_leak_fix.png, /default.tar.gz) since the stated intent is "repo-root build/debug artifacts".

2. AGENTS.md:505–516 — converting the Dev Commands fenced bash block into a numbered plain-text list loses syntax highlighting and one-shot copy/paste of the whole sequence, with no apparent gain over the original — why it matters: AGENTS.md is contributor-facing; the fenced block was the more useful rendering — suggestion: keep the fence and just prepend the "run from ui-tui/" note as a comment or lead-in line.

3. acp_adapter/server.py:~227 — logger.warning added to the HERMES_VERSION import fallback runs at module import time, before logging configuration exists in most entry paths, so it emits via the stdlib last-resort handler straight to stderr — why it matters: in packaging modes where hermes_cli is legitimately absent from the ACP sidecar this becomes recurring stderr noise rather than a diagnosable signal — suggestion: keep the message but route it through the same once-per-process/debug-level pattern used by other import fallbacks in this file, or attach the reason (exc_info=True).

4. Nit (.gitignore:103): *.png.bak is likewise unanchored; if it's meant for editor swap files next to the deleted screenshot, scope it (/*.png.bak) for symmetry with item 1.

Overall: worthwhile housekeeping — dropping committed binaries plus documenting the deliberate shell=True exception in secret_sources/base.py are genuine improvements; no blocking issues found.

— reviewer-a · automated agent review (Hermes week-review)

[teknium1]

Thanks @EAbaracus — the artifact cleanup half of this PR was salvaged into #92394 (your commit cherry-picked with authorship preserved) and is now merged as 0cde4dd, together with .gitignore root-anchoring per the review feedback and matching .dockerignore rules so root archives can't reach Docker image layers via COPY . . again.

One correction to the assessment here: default.tar.gz was not just an unreferenced test byproduct — it was a full profile export whose config.yaml contained a populated webhook secret, so the removal was security-relevant and we fast-tracked it. The credential has been handled separately.

The other two commits (ACP version-import warning, AGENTS.md doc tweaks) had open review concerns (import-time logging before logging config exists; fenced-block regression), so we're not taking them as-is — feel free to reopen those as separate focused PRs if you'd like to iterate. Closing this one since the substantive fix has landed. Appreciate the audit work!