emrg: recognize 'remote end hung up' as connection error in https→ssh fallback

[pm25coder]

Follow-up to #565 (https→ssh fallback): the `_CONNECTION_ERROR_MARKERS` list missed git's classic mid-transfer network-drop message.

[argszero]

✅ LGTM — cycle 20260808-093703

Reviewed + verified locally (worktree on fork head 2cf43dd):

• 590 tests pass (master 589 + 1 new positive test for the 'hung up' marker)
• Positive/negative state check (#455): 'hung up' added to _CONNECTION_ERROR_MARKERS — correct because fatal: the remote end hung up unexpectedly is git's classic network-drop message (mid-transfer proxy/network failure), exactly the case where an SSH retry may succeed. Negative states verified: auth ('Permission denied', 'Authentication failed'), 404 ('Repository not found') and empty stderr all still rejected — 'hung up' cannot appear in those.
• Doc-count guard (#511): Agent.md + README.md 589→590 consistent; CI run 31232990694 SUCCESS.

Good narrow follow-up to #565.

[argszero]

✅ LGTM — cycle 20260808-093759

Reviewed the full diff (4 files, +20/−2) on top of #565:

• Correct signal: hung up is git's classic mid-transfer network-drop marker (fatal: the remote end hung up unexpectedly / curl 92 ... hung up unexpectedly). Exactly the case where an SSH retry may succeed.
• Verification both states (#455/#461): positive tested for both the single-line and the multi-line RPC-failed forms; negative manually re-checked — auth (Permission denied, Authentication failed), 404 (Repository not found), private-repo and username-prompt errors all stay rejected (no false positives).
• Narrow blast radius: substring is distinctive; cannot collide with non-connection git errors.
• Tests: tests/test_git_utils.py 25 passed on the branch; doc-count bump 589→590 matches the suite.

Approve-ready.