Adding a loss operator

A new glyph is a parser arm plus a rule_loss branch plus a catalog page plus a 2-D row. It is not a new adversarial recipe. Pairing, b_cap, feature matching, cover, and the particle cloud stay in conceptmod.toys (locked_adv_defaults, reject_unlocked). Call those. Do not copy GANLoss or GradRegularizer into conceptmod, and do not retune κ, FM, or the demo cover inside the op.

Pages in docs/dsl/operators/ are written by hand. This repo does not generate them from a decorator.

Before you add a glyph

Read docs/dsl.md "Jobs that look missing — already recipes". If the job is already c--, a=b, c++, or c--|k#k, do not add c!!, +, or a keep+erase token. The tests in tests/test_dsl_jobs.py exist to stop that.

Also check the dead list. ; stays unimplemented until there is a real scorer. ^ already has a rule_loss branch; the 2-D gap is the missing render, not a missing parser arm.

Files

File	What changes
conceptmod/dsl.py	One arm in _parse_concept, in the same order the original dispatcher used. A constant next to EXAGGERATE. describe_phrase / _describe_rule get one sentence that matches the loss.
conceptmod/ops.py	One branch in rule_loss. Defaults live on OpDefaults, not as literals scattered through the branch.
tests/test_dsl.py	The string parses, and the strings that must fail still fail.
tests/test_ops.py	The velocity target on a tiny backend, including the default guidance.
docs/dsl/operators/<name>.md	Syntax, velocity target, defaults, composition, 2-D verdict, failure modes, a snippet that calls the live parser.
docs/dsl/operators/index.md	One row. If the op cannot run on TwoAxisBackend, say so.
docs/dsl.md	One scoreboard row after you have actually run the 2-D fixture. right / needs help / recipe / dead. Do not write the verdict first.
conceptmod/game.py	Only if the new op can run on TwoAxisBackend. If it cannot, refuse it in loss_game the way ^ and ; are refused. Do not stub a fake loss.

conceptmod/toys/ is not in this list. A loss op does not get its own gan_mode, κ, or cover weight.

Checklist

1. Parse the glyph in dsl.py. Keep |, :alpha, and :key=value working. New options are floats, same as guidance.
2. Implement rule_loss with the frozen/trained velocity notation already used in ops.py (v_f / v_t, CFG as v(p) − v("")).
3. If the 2-D fixture can express it, add a job in conceptmod/analysis_dsl.py that goes through rule_loss and extend dsl.md from the run. If it cannot, say dead here and why (missing render, missing scorer, uncond embedding is zero). Do not invent a probe that hides the gap.
4. Write the catalog page from the code you just ran, including the default that OpDefaults actually ships.
5. Decide whether loss_game can step it. Refuse with GameError when the fixture cannot. A formulation PASS must not be reachable for a loss that never ran.
6. pytest tests/test_dsl.py tests/test_ops.py tests/test_dsl_jobs.py tests/test_game.py tests/test_dsl_catalog.py -q

GPU proofs (SANA, Z-Image, Anima, Krea) are a separate run. A green 2-D job is not that run. Say so on the page.

What you do not fork

• locked_adv_defaults() — the game and the 2-D suite already call it. A new op inherits that stamp by running under loss_game or under analysis_2d. It does not take an adv= recipe of its own.
• reject_unlocked — stranger pairing, FM-on, and thinned b_cap are already refused when a game claims locked. Do not re-encode those checks inside the op.
• Erase variants GEM / EA — conceptmod/ops_erase.py. They are not phrase glyphs. Wiring a new erase mode means extending that hook and the 2-D method list, not a new token, and still not a new adv stamp.