feat(optimize): add --subtract, the subtraction axis, behind a deterministic floor

Every command so far asked an addition question — what to add, what to move,
what it costs. Nothing asked what is no longer earning its always-loaded rent.
This adds that axis as a fourth lensCheck on the existing hybrid motor rather
than a new scanner or a 22nd command: the measured payoff (~18% of one file)
justifies a mode, not machinery.

It is the only lens that proposes REMOVING config, so it carries a guarantee
the others don't need: a load-bearing block is never a candidate. Precision is
asymmetric — a missed dead line costs a few tokens per turn, a deleted one
costs a wrong remote or a broken script — so the floor is decided in code
(lib/floor-exclusion.mjs) before the opus judge sees anything, never in prose.

Granularity is the leaf block, with two structural exceptions: a paragraph
ending in ':' merges with the list it introduces, and an ordered list is a
contract whose steps inherit floor from any sibling. Unordered lists
deliberately do not inherit — a load-bearing bullet and a disposable one
routinely share a list, and container-reasoning is the error the hand-built
ground truth exists to catch.

Verified against that ground truth (built before any classifier existed), with
the comparison machine-checked rather than read by eye: zero load-bearing
blocks proposed, 11/18 deletable groups surfaced, ~756 tok ~ 18% of a ~4300
token file — inside the pre-registered band. The first run found five floor
violations the synthesized fixture missed; each got a structural rule and a
fixture shape so it cannot regress.

Three real bugs the dogfood run exposed, all now covered:
- JS \b is ASCII-only, so /\bunngå\b/ never matches — every Norwegian keyword
  ending in æ/ø/å was silently dead.
- A bare word/word is not a path; "pros/cons" vetoed the largest deletable
  block until PATH_RE was tightened to rooted paths and globs.
- "Mid-sentence" must key on a preceding lowercase letter; the loose version
  read **bold labels:** and quoted openers as entities, costing 4 of 11 groups.

BP-SUB-001 is grounded entirely in the Anthropic steering blog already cited by
BP-MECH-001..004 and asserts nothing from the talk that motivated the feature —
no "80%", no ablation figure.

Suite 1365 -> 1382/0. Frozen v5.0.0 snapshots untouched; plain optimize output
byte-identical on identical input (--subtract adds keys only when passed).
knowledge-refresh-cli's reference date moved to 2026-08-01: its premise that
every seed entry was verified 2026-06-20 expired when BP-SUB-001 got a genuine
verification date, and backdating the entry to fit the test would have been a
lie about when its source was checked.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RW2haJXbxZpKivKHseSXNh

docs/scanner-internals.md

@@ -499,6 +499,56 @@ orchestrates pre-filter→agent→report. **Agent-driven → deliberately NOT by
 outside the snapshot suite); the pre-filter lib *is* unit-tested (13 tests). No new orchestrated
 scanner → scanner count stays 15; agents 6→7, commands 18→19, suite 1055→1068.
 
+### Subtraction lens (`optimize --subtract`, `BP-SUB-001`) — the inverse axis
+
+**Why a mode, not a command or scanner.** Every other command asks an addition question; nothing
+asked what is no longer earning its rent. A hand-built ground truth over a real 250-line global
+CLAUDE.md put the honest payoff at ~850–1400 always-loaded tokens of ~4300 (~20 %, not the source
+anecdote's 80 %), with 26 of 34 blocks load-bearing. That proportion justifies a fourth `lensCheck`
+on the existing hybrid motor — not a new scanner (no badge bump 16→17, no frozen-snapshot risk) and
+not a 22nd command. Cross-session state is what would have forced a command, and the earned-re-add
+*ledger* is deliberately out of v1.
+
+**Polarity is flipped from `lens-prefilter`.** That module is recall-first because a false candidate
+only costs the judge a moment. Here a false candidate is a proposal to *delete*, so
+`subtraction-prefilter.mjs` is precision-first and carries a blocking deterministic guarantee.
+
+**Floor-exclusion is a separate module on purpose** (`lib/floor-exclusion.mjs`). §6.0's asymmetry —
+a missed dead line costs a few tokens per turn, a deleted load-bearing line costs a wrong remote —
+means the guarantee must not rest on a probabilistic judge, so the floor runs *before* the agent
+and is legible as its own unit. Markers: code span, URL/host, filename, rooted path, version pin,
+policy invariant, and `unresolved-entity` (a mixed-case capitalized word mid-sentence). The last is
+a deliberate conservative default — resolving "Forgejo" from an ordinary capitalized word needs a
+dictionary, so the mechanism declines and keeps the block.
+
+**Granularity: leaf block + two structural exceptions.** (1) A paragraph ending in `:` merges with
+the list it introduces — a stem often carries no literal of its own, and deleting it without its
+list is meaningless. (2) An *ordered* list is a contract: steps inherit floor from any sibling,
+because deleting step 2 of a five-step protocol is not like dropping one platitude. Unordered lists
+do **not** inherit — a load-bearing bullet and a disposable one routinely share a list, and
+container-reasoning is exactly the error the ground truth was built to catch.
+
+**Three lessons from the dogfood run**, all invisible to the synthesized fixture and worth keeping:
+- **JS `\b` is ASCII-only.** `/\bunngå\b/` never matches — the trailing `å` is not a word character,
+  so there is no boundary after it. Every Norwegian keyword ending in æ/ø/å was silently dead. Use
+  the `LB`/`RB` lookaround constants, never `\b`, around that vocabulary.
+- **A bare `word/word` is not a path.** `pros/cons` vetoed the single largest deletable block until
+  `PATH_RE` was tightened to rooted paths and globs; real filenames are `FILENAME_RE`'s job.
+- **"Mid-sentence" must key on a preceding lowercase letter**, not on "anything that is not a full
+  stop". The loose version read `**Bold labels:**` and quoted openers (`"Som AI kan jeg ikke…"`) as
+  entities and cost 4 of 11 deletable groups.
+
+**Measured against the ground truth:** zero load-bearing blocks proposed (the blocking §8 gate),
+11/18 deletable groups surfaced, ≈756 tok ≈ 18 % of the file — inside the pre-registered band. The
+misses are all the conservative default working as designed (entity names, code spans, and
+declarative/infinitive phrasing that carries no imperative). No new scanner: scanners stay 16,
+commands 21, agents 7; suite 1365→1382.
+
+**Note for a future narrowing of the veto:** the two failure modes the agent prompt hardens against
+— staleness-is-not-deletion and tier-2-is-not-tier-3 — are currently *also* covered by exclusion
+(both example blocks carry code spans/version pins and never reach the judge). The prompt language
+is the only protection if those markers are ever loosened.
+
 **Test-isolation fix (this session):** `token-hotspots.test.mjs` `runScanner` now wraps `scan()` in
 the shared `withHermeticHome` helper — the suite is green on BOTH a real and a clean `HOME` (the OPT
 section's old «run with clean HOME» caveat is resolved). Snapshot/byte tests were already hermetic.