revert: unproven default-off features per merge policy (wins only)

[Brian Krabach (bkrabach)]

Policy

Owner-approved merge policy: main carries wins only -- bug fixes with fail-before tests, or
features whose eval PROVED a win. Unproven default-off features live on branches/draft PRs, not
main. This PR reverts six merged features that shipped without a proven win (several explicitly
measured as not a win), while keeping the one proven fix and the one enabling infra change.

Reverted (newest to oldest; each is a standalone git revert -m 1)

• fix: _capture_fork_prefix no longer forks onto an array that was never sent (#jnt) #28 fix: _capture_fork_prefix no longer forks onto an array that was never sent -- not
  in the original revert list, but merged after feat: summary_call_mode -- cache-safe fork of the summarization call (#7k2) #27 and exists solely to fix a bug in the
  summary_call_mode="fork" code path feat: summary_call_mode -- cache-safe fork of the summarization call (#7k2) #27 introduced. Once feat: summary_call_mode -- cache-safe fork of the summarization call (#7k2) #27 is reverted there is no surviving
  target for this fix (it would be dead code / an orphaned test file referencing a feature no
  longer on main), so it is reverted first, ahead of feat: summary_call_mode -- cache-safe fork of the summarization call (#7k2) #27, to keep feat: summary_call_mode -- cache-safe fork of the summarization call (#7k2) #27's own revert clean.
• feat: summary_call_mode -- cache-safe fork of the summarization call (#7k2) #27 feat: summary_call_mode -- cache-safe fork of the summarization call -- merged with
  benefit explicitly labeled unmeasured ("benefit correctly labeled unmeasured" in the
  original merge note).
• feat: clear_at_least -- a worth-the-rebuild predicate in front of compaction (#2o9) #26 feat: clear_at_least -- a worth-the-rebuild predicate in front of compaction --
  falsified: no threshold passes; it cuts boundaries by killing turns rather than helping.
• feat: replay_last_user_on_compaction -- opt-in last-user replay at the tail #24 feat: replay_last_user_on_compaction -- opt-in, unproven.
• feat: token_meter "hybrid" -- provider-anchored counts with provenance (+ the never-run estimate-vs-hybrid-vs-actual measurement) #22 feat: token_meter "hybrid" -- its own estimate-vs-hybrid-vs-actual measurement was
  never run.
• feat: tool-result budget (token-denominated, head+tail, per-tool) + spill-to-disk -- all defaults no-op #21 feat: tool-result budget (token-denominated, head+tail, per-tool) + spill-to-disk --
  unproven, all-defaults-no-op.
• feat: compaction_strategy "summary" — LLM rolling-summary compaction (opt-in) #20 feat: compaction_strategy "summary" -- measured +83% run cost, summarizer share
  only 26-35% of that cost (i.e. the regression isn't even mostly the summarizer call itself);
  gate was not met.

Kept

• fix: protected_tool_results=0 must protect ZERO tool results, not all of them #23 fix: protected_tool_results=0 must protect ZERO tool results, not all of them -- real
  bug fix (inverted-knob-meaning bug at exactly one value), with its own test
  (tests/test_protected_tool_results.py). Verified byte-identical:
  diff <(git show e9a1697:README.md) README.md and the equivalent for the protected_tool_results
  function in __init__.py are both empty after the reverts.
• feat: real-usage token meter (token_meter: actual) for compaction triggering #19 feat: real-usage token meter (token_meter: "actual") -- needed for the proven
  compact-late result; kept in its original, unmodified form (README.md/DONE-NOTE.md diffed
  byte-identical against this PR's HEAD).

Feature branches preserved for later evaluation

The reverted work is not deleted -- it remains evaluable on its original lane branches:

• lane/pmt-fork-span-predicate (open PR feat: span-size predicate in front of the summarizer fork (summary_fork_min_span_ratio / summary_call_mode: auto) #29, builds on feat: summary_call_mode -- cache-safe fork of the summarization call (#7k2) #27/fix: _capture_fork_prefix no longer forks onto an array that was never sent (#jnt) #28 -- left untouched; note it will
  need to be rebased onto this revert or evaluated standalone once feat: summary_call_mode -- cache-safe fork of the summarization call (#7k2) #27/fix: _capture_fork_prefix no longer forks onto an array that was never sent (#jnt) #28 are off main)
• lane/7k2-summary-call-fork (feat: summary_call_mode -- cache-safe fork of the summarization call (#7k2) #27)
• lane/jnt-fork-prefix-capture (fix: _capture_fork_prefix no longer forks onto an array that was never sent (#jnt) #28)
• lane/2o9-clear-at-least (feat: clear_at_least -- a worth-the-rebuild predicate in front of compaction (#2o9) #26)
• lane/l8-replay-last-user (feat: replay_last_user_on_compaction -- opt-in last-user replay at the tail #24)
• lane/q69-token-meter-hybrid (feat: token_meter "hybrid" -- provider-anchored counts with provenance (+ the never-run estimate-vs-hybrid-vs-actual measurement) #22)
• lane/x1r-tool-result-budget (feat: tool-result budget (token-denominated, head+tail, per-tool) + spill-to-disk -- all defaults no-op #21)
• feat/summary-compaction-strategy (feat: compaction_strategy "summary" — LLM rolling-summary compaction (opt-in) #20)

Test result

Full suite green on the revert branch: 87 passed (uv run pytest -q), ruff check . clean.
(Baseline before any revert on this branch: 286 passed, 1 skipped -- matches #28's own merge note.)