Put one canonical exemplar into each surface's generation path (closes #834)

[atomchung]

What this closes

#832's post-merge rerun (receipts on PR #833) bent the length curve for the first
time in three of four frozen scenes, and regressed the fourth — the no-book
freeform one, the surface with the emptiest guidance and no engine payload
anchoring the answer. The diagnosis in that report: the canonical exemplars were
wired into QC (tests/agent/expression-witnesses.json) but never into the
generation path, and an example only steers output when the model sees it while
writing. This puts one there, per surface.

Placement only — #832's stop condition holds. No new rule, no registry ID, no
length cap, no obligation.

The four exemplars

Each conversational surface's reference file now opens with its own copy,
which is where progressive disclosure loads it: when that surface is in play.

Surface	Reference file	Scene	Why this one
consider	references/trade-consequence.md	consider_three_way_comparison	#830 acceptance template 1, owner-approved; the only witness that exercises a multi-candidate stance, a counter-side that exists only as the overturning line, and a full caliber end block. WDGT/GRDC/FABR.
no-book	references/decision-framing.md	no_book_single_name	#830 acceptance template 3, owner-approved; every book-derived claim refused and a stance still landed. This is the shape the regressed live scene was answering in. WDGT.
freeform	references/freeform-answers.md	freeform_positions_view	The structural derivation stays empty (rule 4 is unchanged), so the exemplar is the whole of what this file shows: it is the only freeform witness that exercises all four floors, including the bottom floor as a single offer rather than a section. GRDC.
weekly	references/weekly-market-read.md	weekly_read_connection	The fullest weekly witness: the one engine↔book connection leads, the bound on what the alert claims follows, the next-week check closes. No issuer named.

Every issuer is fictional and nothing is derived from a user record — the corpus
already guaranteed that and the copies are verbatim.

Single source, mechanically

• Marker format: a fenced block whose info string is exemplar <scene_id>.
  One line, no counting, no heading walk — and the id travels with the text
  instead of sitting in a table beside it.
• Sync test: tests/test_expression_contract.py::test_every_surface_reference_opens_with_its_canonical_exemplar
  normalizes both copies (per-line rstrip + strip) and fails closed on drift, so
  editing either side alone reddens.
• The pairing is never hand-listed. Which reference file owns a surface is
  read from the corpus's own surfaces map; which scene a file carries is read
  from that file's own fence. The test compares two declarations that already
  existed rather than adding a third to keep in step.
• Mutation proofs: test_exemplar_drift_is_caught (a one-character edit must
  redden; an untagged fence must stop being read as an exemplar) and
  test_no_other_document_carries_an_exemplar_block (the fence cannot spread to
  a fifth file).

The one interpretation I had to make

tests/test_doc_language.py::test_implementation_markdown_is_english_only covers
every .md under skills/fomo-kernel/, and three of the four exemplars are this
product speaking Traditional Chinese. Pasting them was a hard conflict with that
gate, and the issue's own wording (~300 zh chars, drawn from the existing
witnesses) rules out substituting English ones — there are no English witnesses
for the no-book and freeform surfaces.

Resolution: the gate now skips exactly the lines inside a tagged exemplar
fence in a surface reference, on the ground that quoted product output is not
documentation prose — the same distinction that already keeps copy/zh-TW.json
out of ENGLISH_IMPLEMENTATION_ASSETS. The exemption is bounded three ways at
once: by directory (references/ only), by fence tag, and by the byte-identity
the sync test proves. test_the_exemplar_carve_out_stays_narrow is its mutation
proof — the same Chinese line stays visible in another directory, in an untagged
fence, and in the prose on either side of the fence.

Always-loaded byte budget

File	Before	After
AGENTS.md	7,808	7,808
skills/fomo-kernel/SKILL.md	8,512	8,563
Pair total	16,320 / 16,384	16,371 / 16,384 (13 bytes free)

SKILL.md gained one clause in "Shape of the answer": "Each reference opens with
its exemplar; follow it." — the pointer the issue allows if bytes allow. Nothing
else moved into the always-loaded layer.

Tests

Both offline groups, on python3.12, from a clean worktree off origin/main@6aabbcd:

python3 tests/run_all.py --group product   →  PASS: all 48 suites passed
python3 tests/run_all.py --group qa-eval   →  PASS: all 11 suites passed

The two suites this touches:

• tests/test_expression_contract.py — 24/24 (21 before; +3: the sync test,
  the stray-fence gate, the drift mutation proof)
• tests/test_doc_language.py — 36 (35 before; +1: the carve-out mutation
  proof)

Files changed

• skills/fomo-kernel/references/trade-consequence.md, decision-framing.md,
  freeform-answers.md, weekly-market-read.md — each opens with a
  ## The exemplar section: three lines of what the exemplar demonstrates, the
  provenance line, and the verbatim block.
• skills/fomo-kernel/SKILL.md — one clause pointing at them.
• docs/expression-contract.md §3.5 — records that the corpus is no longer the
  only place the exemplars live, and that the two copies are one text.
• docs/maintainer-guide.md — the [implementation·skill] Put one canonical exemplar into each surface's generation path — examples-first, closing #832's residual force #834 mirrored-surfaces row: exemplar ↔ witness
  sync ownership, the marker format, the English-only carve-out and its bounds.
• tests/test_expression_contract.py — the sync test, the stray-fence gate, and
  the drift mutation proof.
• tests/test_doc_language.py — the scoped English-only carve-out and its
  mutation proof.

Not in this PR

The live proof. Per the issue, the frozen four-scene rerun happens post-merge in
the private harness; the expectation on record is convergence toward exemplar
scale with the freeform regression reversed.

🤖 Generated with Claude Code

[atomchung]

Post-merge live measurement (sanitized): every scene drops; the regression reverses; writes go perfect

Fourth frozen-scene rerun against merged main 94d28d0, same model and inputs as every prior round. First-answer zh chars across the four interventions (loosened → obligation-deletion → mother law → exemplars-in-path):

• cash-deployment: 1,839 → 1,755 → 1,098 → 1,227 (conversation total 5,360 → 2,334)
• no-book single name: 1,495 → 1,524 → 1,849 ⚠️ → 959 — the surface that regressed under the mother law flipped hardest once its reference opened with an exemplar, exactly as diagnosed
• discovery: 2,670 → 2,696 → 2,310 → 1,719
• comparison+selection: 2,297 → 2,204 → 1,888 → 1,356; the recording confirmation runs ~345 chars — template scale

Writes this round: exploration produced zero canonical rows on all surfaces (the discovery scene explicitly told the user its six trials left no record), and the selection scene produced exactly one row for the user-selected candidate. No process-narration openers observed.

Mechanism now demonstrated end-to-end across four rounds: prose rules alone < obligation deletion < positive shape norm < shape norm + a visible exemplar — only layers that reach the generation path move the needle, which is the documented examples-first guidance. Remaining scale (full analyses at ~1.0–1.7k zh chars ≈ 2–3 phone screens) is information-dense rather than obligation discharge; per #832's stop condition no further rules are added — whether to compress further is an owner judgment from real usage.