3-arch QEMU acceptance + live diagnostic logs for the BAM reconciliation fix

Per .claude/CLAUDE.md's non-negotiable acceptance criteria, ran
`make -f Makefile.starkernel ARCH=<amd64|aarch64|riscv64> clean qemu`
sequentially (one instance at a time, TCG) against the preceding commit.
All three booted cleanly to `(zuse) ok>` with the full Tripod fleet
(Hera/Hermes/Artemis) alive and identical dictionary hashes:

  Mama:    dict_hash=0x8d712e07c690d713
  Hermes:  dict_hash=0xa07a57d794c1665d
  Artemis: dict_hash=0x9cc9a7e72b590250

Also includes the amd64 run (20260905-155329) that carried the temporary
TEST-BAM-RECON diagnostic via SK_CMD, confirming the fix live before the
diagnostic word itself was reverted: TEST-BAM-RECON: PASS, no FAIL lines.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018EjXFo7mPXjUMjfJeuUUz4

FABRIC-2.md

@@ -4441,6 +4441,30 @@ stale the way the original carry-forwards did.
   Not fixed in this pass; flagged precisely so it isn't rediscovered as a mystery later.
   *(Same location as above.)*
 
+  **CLOSED 2026-09-05.** Traced against current code before writing anything, per this
+  document's own discipline: `blk_meta_relocate_devblock()` turned out **not** to be part of
+  the gap after all — it calls `blk_subsys_relocate_block()` per member block, which already
+  calls `blk_mark_free()` on the home LBN and `blk_update()` (which sets the BAM bit) on the
+  target, so relocation was already keeping BAM in sync. The real gap was entirely in
+  `blk_firsttouch_claim()`'s side (never touched BAM at all) and in the reverse direction
+  (`devblock_is_free()`/`blk_allocate()` never checked `BLK_FLAG_CLAIMED`), confirmed by a
+  full-documentation audit ([[project_fabric012_full_audit_20260905]] in memory) before
+  fixing. **Fix** (`block_subsystem.c`): `devblock_is_free()` (shared by
+  `blk_firsttouch_claim()` and `blk_migration_idle_check()`) now also checks the BAM entries of
+  all `BLK_PACK_RATIO` member LBNs, not just `blk_meta_t`; a new `devblock_claimed_by_lbn()`
+  helper is wired into `blk_allocate()`'s free-scan so it skips any LBN whose devblock is
+  `BLK_FLAG_CLAIMED`; `blk_firsttouch_claim()` now calls `blk_mark_allocated()` on all 3 member
+  LBNs of each devblock it claims, which also fixes a quiet side effect where
+  `vol_meta.free_blocks` never decremented for FIRSTTOUCH claims. Both boundary edge cases
+  (the reserved/user LBN split at a slot's own `start_lbn`, and BAM-array bounds) are guarded
+  explicitly rather than assumed safe. **Verified, not just compiled clean**: full 3-arch QEMU
+  acceptance (identical dictionary hashes, clean `BYE`, no regression) plus a live logic test —
+  a temporary `TEST-BAM-RECON` word (written, run once via `SK_CMD`, then fully reverted,
+  per this project's own probe discipline) ordinary-allocated 4 LBNs, `FIRSTTOUCH`-claimed 2
+  devblocks, and confirmed on real running code that neither direction of the gap reproduces:
+  the ordinary allocations and the claim landed on disjoint LBN ranges, and `blk_allocate()`
+  never returned an LBN from the claimed range afterward. `PASS`, no `FAIL` lines.
+
 ### I.3 — Milestone 5 remainder (BINDSTEP) — CLOSED 2026-09-04
 
 - [x] **`EJECT` / detach behavior.** Scoped 2026-08-27 (§F.10): flush via

src/block_subsystem.c

@@ -1078,6 +1078,34 @@ int blk_mark_free(uint32_t block_num) {
     return BLK_OK;
 }
 
+/* blk_meta_t/BAM reconciliation, 2026-09-05 (closes the other half of the
+ * gap §I.2 flagged): blk_allocate()'s BAM scan below must not hand out an
+ * LBN whose devblock a FIRSTTOUCH claim already owns. BMAPFMT ownership
+ * lives only on the devblock's pack_offset==0 "representative" LBN's own
+ * blk_meta_t (blk_firsttouch_claim() never stamps the other two members),
+ * so an arbitrary candidate lbn must walk back to its representative
+ * before checking BLK_FLAG_CLAIMED.
+ *
+ * Reserved/user boundary case: lbn_to_slot_pbn() adds the fixed
+ * BLK_DISK_SYS_RESERVED (32) offset before computing pack_offset, so a
+ * slot's own start_lbn is not guaranteed to land on pack_offset==0 --
+ * walking back by pack_offset can land before slot->start_lbn entirely
+ * (the representative sits in the reserved system region, which has no
+ * LBN and thus no blk_meta_t at all). FIRSTTOUCH's own scan already
+ * skips any lbn where slot_pbn_pack_offset(rel_pbn) != 0, so it can never
+ * reach that representative to claim it -- such an lbn is guaranteed
+ * unclaimed; returning 0 here without calling blk_get_meta() on an
+ * address outside the slot is both correct and the safe choice. */
+static int devblock_claimed_by_lbn(blk_dev_slot_t *slot, uint32_t lbn) {
+    uint32_t rel_pbn = lbn_to_slot_pbn(slot, lbn);
+    uint32_t pack    = slot_pbn_pack_offset(rel_pbn);
+    uint32_t offset  = lbn - slot->start_lbn;
+    if (offset < pack) return 0;
+    blk_meta_t meta;
+    if (blk_get_meta(lbn - pack, &meta) != BLK_OK) return 0;
+    return (meta.flags & BLK_FLAG_CLAIMED) != 0;
+}
+
 int blk_allocate(uint32_t *block_num) {
     if (!block_num)      return BLK_EINVAL;
     if (!g.initialized)  return BLK_ENODEV;
@@ -1094,7 +1122,7 @@ int blk_allocate(uint32_t *block_num) {
 
         for (uint32_t i = 0; i < limit; i++) {
             uint32_t k = (hint + i) % limit;
-            if (!s->bam[k].allocated) {
+            if (!s->bam[k].allocated && !devblock_claimed_by_lbn(s, s->start_lbn + k)) {
                 s->bam[k].allocated = 1;
                 s->bam_dirty        = 1;
                 if (s->vol_meta.free_blocks) s->vol_meta.free_blocks--;
@@ -1360,10 +1388,31 @@ int blk_flags_set(uint32_t block_num, uint64_t flags) {
  * misclassify a claimed-but-not-yet-fp-stamped or fp-stamped-but-
  * evicted devblock, neither of which should exist given owner_fp is
  * always stamped/cleared alongside the flag below, but checking both is
- * one extra memcmp for a real safety margin, not paranoia without cost. */
-static int devblock_is_free(const blk_meta_t *m) {
+ * one extra memcmp for a real safety margin, not paranoia without cost.
+ *
+ * blk_meta_t/BAM reconciliation, 2026-09-05 (closes the gap §I.2 itself
+ * flagged, not fixed): a devblock can be "unowned" by blk_meta_t and
+ * still be in active use via the generic BAM (ordinary BLOCK/UPDATE
+ * traffic, unrelated to BMAPFMT ownership) -- claiming it anyway would
+ * silently corrupt that content. Checks all BLK_PACK_RATIO member LBNs,
+ * not just rep_lbn itself, since BAM is tracked per-1KiB-block while
+ * BMAPFMT ownership is tracked per-devblock. rep_lbn is always the
+ * pack_offset==0 member (both call sites already filter to that), so
+ * rep_lbn+1/+2 are its devblock siblings. Bounds-checked against
+ * slot->user_blocks rather than assumed divisible by BLK_PACK_RATIO --
+ * the very first user LBN in a slot can belong to a devblock whose other
+ * members fall below start_lbn entirely (see the reserved/user boundary
+ * note in blk_allocate()'s new claimed-check below). */
+static int devblock_is_free(blk_dev_slot_t *slot, uint32_t rep_lbn, const blk_meta_t *m) {
     static const uint8_t zero_fp[8] = {0};
-    return !(m->flags & BLK_FLAG_CLAIMED) && memcmp(m->owner_fp, zero_fp, 8) == 0;
+    if ((m->flags & BLK_FLAG_CLAIMED) || memcmp(m->owner_fp, zero_fp, 8) != 0)
+        return 0;
+    for (uint32_t i = 0; i < BLK_PACK_RATIO; i++) {
+        uint32_t off = (rep_lbn - slot->start_lbn) + i;
+        if (off >= slot->user_blocks) return 0;
+        if (slot->bam[off].allocated) return 0;
+    }
+    return 1;
 }
 
 /* Sane upper bound on one claim, independent of how large count is --
@@ -1398,7 +1447,7 @@ int blk_firsttouch_claim(const uint8_t owner_fp[8], uint32_t count, uint32_t *ou
 
         blk_meta_t meta;
         if (blk_get_meta(lbn, &meta) != BLK_OK) continue;
-        if (devblock_is_free(&meta)) found[nfound++] = lbn;
+        if (devblock_is_free(slot, lbn, &meta)) found[nfound++] = lbn;
     }
 
     if (nfound < count) return BLK_ENOSPC; /* fail outright, no partial claim (§F.11 decision 3) */
@@ -1415,6 +1464,17 @@ int blk_firsttouch_claim(const uint8_t owner_fp[8], uint32_t count, uint32_t *ou
         meta.next_block    = (i + 1 == nfound) ? 0 : found[i + 1];
         meta.chain_length  = nfound;
         if (blk_set_meta(found[i], &meta) != BLK_OK) return BLK_EIO;
+
+        /* blk_meta_t/BAM reconciliation, 2026-09-05: blk_set_meta() above
+         * only touches blk_meta_t -- mark the BAM allocated too, for all
+         * BLK_PACK_RATIO member LBNs of this devblock (not just found[i]
+         * itself), so blk_allocate() can never hand this devblock's
+         * content out from under this claim, and so vol_meta.free_blocks
+         * (blk_mark_allocated()'s own bookkeeping) correctly reflects the
+         * claim instead of silently overstating free space. */
+        for (uint32_t j = 0; j < BLK_PACK_RATIO; j++) {
+            (void) blk_mark_allocated(found[i] + j);
+        }
     }
 
     *out_chain_head = found[0];
@@ -1504,7 +1564,7 @@ void blk_migration_idle_check(void) {
             hot_lbn = lbn;
             have_hot = 1;
         }
-        if (!have_free && devblock_is_free(&meta)) {
+        if (!have_free && devblock_is_free(slot, lbn, &meta)) {
             free_lbn = lbn;
             have_free = 1;
         }