feat(adversary): stateless sign-ascent SRC_STEP for persistent-PGD sources (SPEC §6)

[ocg-goodfire]

What

Implements the sign SRC_STEP variant already named as a valid instantiation in SPEC §6: sources += lr·sign(grad) with the same S15 clamp projection, selected via optimizer: {type: sign, lr_schedule: ...} (a discriminated union with adam on PersistentPGDReconLossConfig.optimizer).

The variant is stateless: PersistentAdversary.opt_state is now SourcesAdamState | None (None iff sign), so the source Adam m/v buffers never exist. On full32L that is −19 GiB/GPU resident (2× the 9.5 GiB source bundle); combined with the validated source_dtype: bfloat16 (#938) the source bundle goes 28.5 → ~4.75 GiB.

Invariants (SPEC IDs)

• §6 SRC_STEP: sign was already a named valid instantiation — no SPEC amendment. Production stays ★ adam.
• S13'/S14': lifecycle unchanged — n_warmup supplemental ascents + one final ascent from the main backward, coeff-unscaled; sources_ascend_project dispatches on the optimizer config.
• S15: same clamp-to-[0,1] after every ascent.
• N1 (fp32 SRC_STEP moments): vacuous for sign — no moments. The adam path (sources_adam_ascend_project) is bit-untouched.

Configs

Four Tier-1 quality A/B arms (pile-4L PPGD-bsc, 10k steps, same seed, dp 8): adam control + sign at lr {0.003, 0.01, 0.03} — a step-size sweep since Adam's normalized per-update magnitude ≈ lr was tuned at 0.01, but sign steps are exactly lr. All arms fp32 source to isolate the update rule. Results will be logged on the bridge task (lp-sign-ascent-source).

Validation

• pytest param_decomp/tests/ green at default device count; adversary/checkpoint tests also green at XLA_FLAGS=--xla_force_host_platform_device_count=4.
• New: test_sign_src_step.py (unit: ±lr move, clamp, dtype, statelessness) + test_step_trains_with_sign_src_step (full jitted train step: losses finite, sources move and stay projected, opt_state stays None).
• make check clean (ruff + basedpyright 0 errors).

🤖 Generated with Claude Code

https://claude.ai/code/session_01ALaUCqhksKmWNbxjfmuMUi

[ocg-goodfire]

Tier-1 quality A/B result: sign-ascent FAILS the gate at all three step sizes (pile-4L PPGD-bsc, 10k steps, same seed, fp32 sources; runs p-e6f2e806 adam / p-992df179 lr.003 / p-41cee32f lr.01 / p-07b7e68b lr.03).

The adversary-independent robustness probe (fresh-PGD 20-step eval) vs adam control:

• lr 0.003: +400–900% throughout — sign adversary far too weak, decomposition never faces real pressure (its better masked CE/KL is the weak-adversary confound, not a win)
• lr 0.01: metastable — oscillates between parity and +300–620% blowups, widening over training (adversary falls behind the co-evolving decomposition)
• lr 0.03: +15–27% consistently, plus +10–25% masked CE/KL and +15–30% CI-L0 — dominated on every axis

Conclusion: Adam's adaptive per-coordinate scaling is load-bearing for the persistent adversary (3 ascents/step tracking a moving target — unlike fresh PGD's 20 dedicated sign steps from scratch). The −19 GiB m/v saving is not available via naive sign; bf16 source+moments (#938, −14.25 GiB) remains the validated path.

Full writeup: lore 2026-07-03--sign-ascent-ppgd-source-fails-quality-gate.

Merge-or-close is the team's call: the impl is a faithful SPEC §6 sign instantiation with tests, but there is now no production use for it. Closing is consistent with the no-dead-code norm; merging keeps the variation point exercised for future signum-style experiments.