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

agents/optimization-lens-agent.md

@@ -33,6 +33,39 @@ whether the line is *really* that kind of instruction:
 | `claude-md-lifecycle-phrasing` | BP-MECH-001 | a recurring automation the model is *told* to perform ("after every commit, run X") — something that should happen deterministically, not at the model's discretion | a **hook** (PreToolUse / PostToolUse / Stop) |
 | `unscoped-path-specific-instruction` | BP-MECH-002 | a constraint that only applies when a *specific* file/path/glob is touched, sitting in root CLAUDE.md where it loads every turn regardless | a **path-scoped rule** (`.claude/rules/` with `paths:` frontmatter) |
 | `never-instruction` | BP-MECH-004 | an *absolute* prohibition — something that must NEVER happen, where relying on the model to remember is the wrong guarantee | a **permission deny rule** or PreToolUse hook |
+| `compensatory-instruction` | BP-SUB-001 | **`--subtract` mode only.** an instruction that corrects general model *behaviour* rather than stating a local fact — so it pays an always-loaded token cost without telling the model anything it could not work out | **removal**, re-added only if the model actually stumbles |
+
+## The subtraction lens (`--subtract` only)
+
+Present only when the payload has a `subtract` block. It asks the inverse of
+every other lens: *what is no longer earning its always-loaded rent?* Three
+things make it different, and all three are non-negotiable.
+
+**1. The floor is not yours to decide.** A deterministic pre-step has already
+excluded every block carrying a local fact — a code span, path, domain, version
+pin, policy invariant, or an unresolved capitalized entity — plus the steps of
+any ordered list whose siblings carry one. You never see those blocks, and you
+must not reason about whether some *other* block ought to be deleted. Judge only
+what you are given. Precision is asymmetric: a missed dead line costs a few
+tokens per turn; a deleted load-bearing line costs a wrong remote, a broken
+script, or a lost afternoon.
+
+**2. Staleness is NOT a deletion signal.** A block that pins an outdated version
+("use Opus 4.8") is a *dead-reference* problem for `drift` / `CA-CML`, not a
+subtraction finding. The instruction is still load-bearing — it encodes a
+decision only the operator can make; it is merely out of date. Recommending
+deletion because content looks stale is a category error. Say "this looks
+outdated" if you must, but never as a removal candidate.
+
+**3. Tier 2 is not tier 3.** Deletable splits into *earned* (compensatory, but
+this model still stumbles on it, so it returns) and *dead* (never missed). Sort
+every candidate into one of the two and say which. A block that has visibly
+earned its place — its subject matter recurs in the repo's own history — is tier
+2 even when its classification is "compensatory". Reporting it as dead weight is
+wrong even though the label matches.
+
+Rank kept candidates by always-loaded token cost, and state the total payoff.
+Frame it as *rent*, never as a mistake: this config was correct when written.
 
 ## Input
 
@@ -45,6 +78,10 @@ You receive an `optimize-lens` payload (JSON) with:
   `recommendation`, `severity`, `source`). Only CONFIRMED register rules reach
   you.
 - `register` — the full confirmed prose-judgment entries, for reference.
+- `subtract` — **present only under `--subtract`.** `{ enabled, candidates,
+  register, detectors }`. Each candidate spans `line`–`endLine` (a whole leaf
+  block, not one line) and carries `signalText` plus the BP-SUB-001 register
+  block. Everything load-bearing was already removed before you saw this.
 
 Always **Read the actual CLAUDE.md file(s)** named in the candidates before
 judging — `signalText` is one line out of context; the surrounding lines decide
@@ -104,6 +141,15 @@ Write `optimization-lens-report.md` to the session directory (≤120 lines).
 {Brief, honest: candidates you dropped and why — "line 22 mentions a path but is
 a cross-reference, not an instruction." This is the precision gate showing its
 work. Keep to a few lines.}
+
+## No longer earning its rent (--subtract only)
+
+{Omit entirely unless the payload has a `subtract` block. Two sub-lists —
+**Dead** (tier 3, out and never missed) and **Earned** (tier 2, out but likely
+to return) — ranked by token cost, with a payoff total. For each:}
+**{file}:{line}-{endLine}** — {what the block says, in one line}  ·  ~{N} tok/turn
+  Tier: {dead | earned — and why}
+  Source: {register.source.url}
 ```
 
 Omit any section with zero kept findings (except keep the "left alone" note when