feat: add inject_mode (suffix/prefix) to TI pipeline

Observation: n4_baseline loss barely moved (1.025→0.965 over 3000 steps),
token_norm grew linearly without plateau — generator likely ignores last-K
CLIP positions (EOS/padding zone) where suffix injects.

Fix: add inject_mode parameter throughout the pipeline:
- "suffix": replace last K positions (original behavior, model may ignore)
- "prefix": replace positions 1:1+K right after BOS — highest attention
  weight in CLIP, much stronger gradient signal expected

Changes:
- selva_textual_inversion_trainer.py: _inject_tokens() helper centralises
  the torch.cat construction for both modes; used in training loop and eval;
  inject_mode stored in checkpoint files
- selva_textual_inversion_loader.py: reads inject_mode from checkpoint,
  includes in TEXTUAL_INVERSION bundle
- selva_sampler.py: uses _inject_tokens() via bundle's inject_mode field
- selva_ti_scheduler.py: inject_mode in _PARAM_DEFAULTS, config, and
  _train_inner call
- ti_sweep_1.json: updated with prefix_inject group (n4, n8, n4+warm);
  n4_baseline marked completed; suffix experiments retained for comparison

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

experiments/ti_sweep_1.json

@@ -1,6 +1,6 @@
 {
   "name": "ti_sweep_1",
-  "description": "First TI sweep: token count, learning rate, and warm init. All generator weights frozen throughout. Baseline = n_tokens=4, lr=1e-3, random init. Primary goal: find a working (n_tokens, lr) pair before optimising further.",
+  "description": "First TI sweep: inject position, token count, learning rate, and warm init. n4_baseline already completed (suffix, loss barely moved — model likely ignores last-K positions). Priority: prefix injection group.",
   "data_dir": "/media/unraid/davinci/Selva/BJ/features",
   "output_root": "/media/unraid/davinci/Selva/BJ/experiment/ti_sweep_1",
   "base": {
@@ -11,52 +11,61 @@
     "seed": 42,
     "init_text": "",
     "lr": 1e-3,
-    "n_tokens": 4
+    "n_tokens": 4,
+    "inject_mode": "suffix"
   },
   "experiments": [
 
     {
       "id": "n4_baseline",
-      "group": "token_count",
-      "description": "4 tokens, lr=1e-3, random init. Primary reference point — all other experiments are measured against this."
+      "group": "suffix_token_count",
+      "description": "4 tokens, suffix, lr=1e-3, random init. COMPLETED — loss 1.025→0.965, nearly flat. Token norm grew linearly to 3.2 with no plateau. Model appears to ignore last-K positions."
     },
     {
       "id": "n8",
-      "group": "token_count",
-      "description": "8 tokens, lr=1e-3, random init. Double the capacity — does it capture more style or just overfit faster?",
+      "group": "suffix_token_count",
+      "description": "8 tokens, suffix, lr=1e-3. More capacity — does it do better than n4_baseline?",
       "n_tokens": 8
     },
+
     {
-      "id": "n16",
-      "group": "token_count",
-      "description": "16 tokens, lr=1e-3, random init. Maximum expressiveness — worth the extra convergence difficulty?",
-      "n_tokens": 16
+      "id": "n4_prefix",
+      "group": "prefix_inject",
+      "description": "4 tokens at positions 1:5 (after BOS). Prefix positions carry the highest attention weight in CLIP — should produce much stronger loss signal than suffix.",
+      "inject_mode": "prefix"
+    },
+    {
+      "id": "n8_prefix",
+      "group": "prefix_inject",
+      "description": "8 tokens at prefix positions. More capacity + high-attention positions.",
+      "n_tokens": 8,
+      "inject_mode": "prefix"
+    },
+    {
+      "id": "n4_prefix_warm",
+      "group": "prefix_inject",
+      "description": "4 tokens, prefix, warm-started from 'mechanical impact sound design'. Best of both: semantically meaningful start + strong gradient signal.",
+      "inject_mode": "prefix",
+      "init_text": "mechanical impact sound design"
     },
 
     {
       "id": "lr_5e4",
       "group": "learning_rate",
-      "description": "n_tokens=4, lr=5e-4. Half the default LR — smoother convergence, possibly better generalisation.",
+      "description": "4 tokens, suffix, lr=5e-4. Slower convergence — mainly a baseline comparison for the prefix group.",
       "lr": 5e-4
     },
     {
       "id": "lr_2e3",
       "group": "learning_rate",
-      "description": "n_tokens=4, lr=2e-3. Double the default LR — faster early convergence, risk of oscillation.",
+      "description": "4 tokens, suffix, lr=2e-3. Faster early movement — does token norm plateau earlier?",
       "lr": 2e-3
     },
 
     {
       "id": "n4_warm",
       "group": "warm_init",
-      "description": "4 tokens warm-started from 'mechanical impact sound design'. CLIP embedding initialises tokens in a semantically relevant region of the space — may converge faster and to a better style representation.",
-      "init_text": "mechanical impact sound design"
-    },
-    {
-      "id": "n8_warm",
-      "group": "warm_init",
-      "description": "8 tokens warm-started from 'mechanical impact sound design'. Combines the warm-init advantage with more expressive capacity.",
-      "n_tokens": 8,
+      "description": "4 tokens, suffix, warm-started from 'mechanical impact sound design'.",
       "init_text": "mechanical impact sound design"
     }