docs: DONE-NOTE for rb1 -- merge-queue repair (rebased and landed #21 and #24)

[Brian Krabach (bkrabach)]

What this is

Merge-queue repair, lane model_performance-rb1. Docs only — this branch's
own diff is a single append to DONE-NOTE.md. The actual work of the lane was
rebasing and landing two other PRs, which are already on main:

PR	subject	landed
#21	tool-result budget (token-denominated, head+tail, per-tool) + spill	49e2799
#24	replay_last_user_on_compaction — opt-in last-user replay at the tail	3972070

routing-matrix #49 was deliberately not touched — a live lane owns that repo.

Suite

point	tree	result
baseline before the lane	f47c894	139 passed
after #21	49e2799	187 passed
after #24 (final main)	3972070	208 passed, 1 skipped
this branch	56270a1	208 passed, 1 skipped

The 1 skip is tests/test_replay_last_user.py:682 — an environment-conditional
skip (amplifier-foundation requires amplifier-core>=1.0.10; this repo pins
core <1.0.10), with the same contract covered unconditionally by a companion
test. It is not a failure. See the deviation note below.

Conflicts resolved (detail in DONE-NOTE.md §1–§2)

• feat: tool-result budget (token-denominated, head+tail, per-tool) + spill-to-disk -- all defaults no-op #21 had two conflicts, not the one predicted: a real code conflict in
  clear() (ours: _reset_hybrid_meter_state() from feat: token_meter "hybrid" -- provider-anchored counts with provenance (+ the never-run estimate-vs-hybrid-vs-actual measurement) #22; theirs: the per-tool
  name map + spill memo resets — diff3 base empty, both kept) and the
  DONE-NOTE.md add/add against fix: protected_tool_results=0 must protect ZERO tool results, not all of them #23's note.
• DONE-NOTE.md is now explicitly a shared, append-only file with a
  convention header and an index table. Both lanes' notes are kept verbatim under
  their original headings. Future add/add conflicts resolve by appending.
• feat: replay_last_user_on_compaction -- opt-in last-user replay at the tail #24 had 8 hunks (7 in __init__.py, 1 in README.md), every one
  additive-vs-additive. Resolution was mechanical and checked: the diff3 base
  section was asserted empty for each hunk before emitting ours-then-theirs.
  git diff origin/main..pr24 = 1,062 insertions, 0 deletions — proof the
  rebase re-applied feat: replay_last_user_on_compaction -- opt-in last-user replay at the tail #24 without displacing anything already on main.

Defaults verified to coexist on merged main

replay_last_user_on_compaction=False · token_meter=TOKEN_METER_ESTIMATE ·
tool_result_budget_tokens=None · compaction_strategy=COMPACTION_STRATEGY_PROGRESSIVE.

Deviation, recorded not smoothed

The brief said #24 was "124/125 — find and fix the one failure". Measured on its
pre-rebase branch: 124 passed, 1 skipped — pass/skip, never pass/fail. There
was no failing test at any point. Making the skip run would require adding a
dependency this repo pins against, which is a dependency change, not a test fix.

model_performance-57p

The merge-queue blocker is cleared — the treatment is on main, default-off.
57p's own blocker (S7 existing and demonstrably discriminating) is unchanged
by this lane. Two things the eval lane must redo because main moved: the
recorded T0 sha256 is stale (re-baseline against 3972070), and #21's five new
default-no-op knobs must also be held at defaults for one-variable discipline.

Spend

$0.00. No API calls, no eval runs, no DTU, no infrastructure created or
registered. All local git / pytest / gh.