fix(memory): keep loop rules resident and trim whole entries (#1443)

[ozand]

Summary

Resolves #1443 by replacing positional tail slicing of memory/index.md with a structural resident/remainder policy:

• The memory index preamble, ## Facts heading, and the five operational rules entries remain resident: Identity, Write target, DO NOT touch, Rules, and Key paths.
• Remaining index lines are selected as complete entries from newest to oldest and restored in source order. No byte/character tail slice can cut a token or remove the resident rules block.
• Memory-fit diagnostics are carried into ContextBuilder.last_fit["memory_index"], so dropped-entry count and byte counts are visible alongside the fix(context): degrade the prompt visibly instead of failing the cycle (#1438) #1440 prompt-fit ladder.
• max_chars is unchanged at 4000.
• The instance memory/index.md is not modified.

Live-size accounting (addendum)

Measured read-only on the live host against the 8,680-byte memory/index.md, with max_chars=4000:

resident rules block:       371 bytes
remainder before:          8,309 bytes
remainder kept after:     3,614 bytes
new memory context:        3,985 bytes
old tail context:          4,000 bytes
net change:                  -15 bytes
whole entries dropped:        28

Thus the resident rules are restored without increasing the memory section; the new result is 15 bytes smaller than the broken tail result. This is compatible with the measured 461-character prompt slack (23,539 / 24,000, 98.08%) and does not invoke the #1440 degradation ladder merely because the rules are restored.

Interaction with #1440

#1440 continues to operate on the assembled sections after get_memory_context. This change makes the memory section structurally safe before that ladder runs and adds memory_index accounting to the existing fit record. The ladder remains responsible for total prompt overflow; the memory reader no longer silently loses the front of the index or reports its post-truncation size as if no content had been dropped.

Tests

• Added regression coverage for the discriminating shape: rules at the head, filler at the tail, small cap; all resident rules survive and no partial marker appears.
• Added context-fit coverage asserting dropped memory entries are visible in last_fit["memory_index"].
• Focused validation: 31 passed (test_memory_v2.py, test_context_prompt_fit.py, test_system_prompt_sections.py, test_log_format_style.py).
• Full pytest: 3582 passed, 18 skipped, 4 failed; the four failures are exactly the required Windows baseline in test_bridge_cycle_tags.py and test_bridge_locking.py.
• Clean origin/main baseline check: tests/test_memory_v2.py passed before the new regression was applied.

Files changed

• nanobot/agent/memory.py
• nanobot/agent/context.py
• tests/test_memory_v2.py

An unrelated upstream #1441 log-format guard was restored in the branch after the milestone WIP history; no #1441 production/test behavior is part of this PR diff against current origin/main.

[ozand]

Reviewed and accepted; one commit added on top (86f937c).

The byte accounting is the right deliverable and it holds — I re-ran the policy against the live 8,680-byte index independently: kept_chars=3985, dropped_entries=28, net −15 bytes against the old tail. All five resident labels present in the output.

What I added, and why. The resident block is selected by matching label text in memory/index.md — an artifact the instance owns and can rename. A rename silently returns that rule to the droppable remainder, which is exactly the failure this policy exists to prevent, and nothing would have said so: an unmatched label reads identically to a matched one. That is the same class as the line-number and glob keys that failed open in #1321 and #1285.

last_index_fit now carries resident_matched / resident_expected / resident_missing, with a warning naming the missing labels. Reported, never asserted — the memory reader must not fail a cycle over index text it does not own. Tests pin both directions: a renamed [DO NOT touch] reports 4 of 5 with the label named, and an intact index reports 5 of 5 so the counter cannot pass vacuously.

One edge left open deliberately. How each session should go (facts/session-flow.md) is still dropped — I verified it is absent from the live-index output. It appeared in #1443's evidence list of entries falling off, but the Proposed direction section named five categories and this PR implements exactly those. Expanding the resident set is a policy call with a permanent cost (every added entry becomes un-droppable), so I am not making it silently. Noted on the issue instead.

Focused suites after my commit: 34 passed.