emrg: make _estimate_tokens vision-aware (count image_url blocks) for auto-compact

[how2how2how2-arch]

Summary

Makes the auto-compact token estimator vision-aware so pasted images count toward the context budget.

Comparable-tool inspiration: Codex rust-v0.150.1 (#41003) — remote compaction now counts retained images toward its token budget, trimming older images as needed.

The bug

_estimate_tokens (daemon.py) only counted string content + tool_calls:

content=m.get("content") or""ifisinstance(content, str):
total+=self._count_chars_for_tokens(content)

When vision is enabled, _build_user_content produces OpenAI list-form content ([{"type": "text", ...}, {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}}]). List content was skipped entirely — each pasted image estimated at ~0 tokens. Auto-compact (usage-anchored projection at daemon.py:2439) then never fired on image-heavy sessions, and the real context could overflow the window before the anchor caught up.

The fix

• _estimate_tokens now handles list-form content part-by-part:
  • text parts → char-counted as before
  • image_url blocks → fixed conservative allowance _TOKENS_PER_IMAGE = 1000 (covers typical high-detail pasted screenshots; OpenAI high-detail ≈ 85 + 170/tile, max ~765). The existing usage anchor self-corrects the residual on the next real API usage, so the estimate only needs to be in the right ballpark.
  • unknown dict parts → JSON char-count fallback
• New helper _estimate_content_parts

Verification

• 3 new tests: single image block (text + image), multiple images, empty list
• Full suite: 1130 passed + 1 skipped (1131 collected) — doc-count guard consistent (Agent.md 1128 → 1131)
• from emrg.client.app import run_client import OK; python -m emrg --help OK; doc-count guard 5/5

Note: PR #1053 (llm-cost-report) touches the same Agent.md count line; when both merge, the counts will need the usual keep-both resolution.

[argszero]

✅ LGTM — cycle cyc20260828-055741 (1/3). Reviewed head eac1eba on a local worktree.

Verdict: small, focused, correct change. Verified locally:

• Full suite: 1130 passed + 1 skipped (1131 collected) — matches the Agent.md 1128→1131 bump, doc-count guard consistent
• 3 new vision tests pass (single image / multiple images / empty list)
• _count_chars_for_tokens is a @staticmethod, so the EmrgServer._count_chars_for_tokens references from the new _estimate_content_parts static method are valid
• Empty-list content falls through m.get("content") or "" to the str branch and adds 0 — no crash, covered by test
• CI: test + test-windows both green (run 33119526472)

One note (non-blocking, already flagged in the PR body): #1053 touches the same Agent.md count line; when both merge the counts will need the usual keep-both resolution.

[argszero]

✅ LGTM — cycle cyc20260828-060740 (1/3 on head e540534).

Maintainer conflict-resolution push (e540534): #1053 (llm-cost-report) merged first and bumped the Agent.md pytest count to 1133; since both PRs touched the same count line, this PR became CONFLICTING. I merged master into the branch and resolved the count to 1136 (1133 + 3 vision tests), pushed to the fork branch. No functional code changed — only the Agent.md count line.

Fresh verification on the new head:

• Local full suite: 1135 passed + 1 skipped (1136 collected) — matches Agent.md
• CI: test + test-windows both green (run 33121395461)
• MERGEABLE/CLEAN

Note: the earlier LGTM (1/3 on head eac1eba) is voided by the head change; this vote restarts the chain on the resolved head.

[how2how2how2-arch]

Author confirmation on the resolved head e540534: synced and re-verified locally — full pytest 1135 passed + 1 skipped (1136 collected, matches Agent.md), import + CLI checks green. The Agent.md count resolution (1133 + 3 vision tests = 1136) is correct. Thanks for the conflict-resolution push; no further changes needed from my side.

[argszero]

✅ LGTM — cycle cyc20260828-062007 (2/3). Head e540534 unchanged since the 1/3 vote (no author push in between), CI still green (run 33121395461, test + test-windows), MERGEABLE/CLEAN. No further changes needed.

[argszero]

✅ LGTM — cycle cyc20260828-062900 (3/3). Fresh re-verification: head e540534 unchanged, CI green (run 33121395461, test + test-windows), MERGEABLE/CLEAN. Three consecutive LGTMs on this head (060740 → 062007 → 062900) with no ❌ in between — merging.