perf(diagnosis): O(n) context truncation instead of O(n^2) (#70)

[JumpTechCode]

Summary

Closes #70.

truncate_for_budget dropped low-priority context items one at a time, and after every single pop re-serialized the entire (shrinking) blob via json.dumps(blob) to re-estimate tokens — quadratic on exactly the large-context case truncation exists to handle.

Fix

Track a running character total: serialize the blob once for a baseline, then subtract len(", ") + len(json.dumps(popped)) per pop. This is byte-exact because popping the last element of a default-json.dumps list removes precisely its leading ", " item separator plus its own serialization, so the running total stays identical to len(json.dumps(blob)).

Why it's safe

• The loop only pops when len(items) > 1 and always pops the last element — which always has a leading separator (never list index 0).
• Default json.dumps serializes any element identically standalone vs nested (no indent, default separators, ensure_ascii=True).
• Verified against a brute-force re-dump over 200 randomized trials (zero mismatches in both dropped-stats and final serialized context), including float / unicode / 2→1-boundary cases.

Behavior is bit-identical to the old per-pop path; only the cost changes.

Test

test_whole_blob_serialized_at_most_once monkeypatches json.dumps and asserts the whole blob is serialized ≤1 time regardless of how many items are dropped (fails on old code at ~120×).

492 unit tests pass; ruff + mypy --strict clean.