fix: give log-node and archive-node real health-check margin

services/archive-node/fly.toml

@@ -32,6 +32,16 @@ primary_region = 'iad'
     type = 'http'
     interval = '15s'
     timeout = '2s'
-    grace_period = '5s'
+    # grace_period MUST exceed boot-to-listening. Measured 2026-07-29 from the
+    # Fly logs: "Starting init" 08:57:14 -> "atrib-archive listening" 08:57:18,
+    # so 4 seconds against a 5 second grace. One second of margin is not margin.
+    #
+    # Startup replays the archived record bodies ("replayed 13 archived
+    # records"), so it grows with opt-in submissions. 13 is small today, which
+    # is exactly why the tight value has not bitten yet.
+    #
+    # 30s is ~7x the measured boot. Fly clamps this field at 60s, so that is the
+    # ceiling; past it the fix is a cheaper startup, not a longer grace.
+    grace_period = '30s'
     method = 'GET'
     path = '/v1/retention'

services/directory-node/fly.toml

@@ -40,4 +40,12 @@ primary_region = 'iad'
     path = '/v6/anchor'
     interval = '30s'
     timeout = '5s'
+    # Left at 15s: measured 2026-07-29, "Starting init" 08:56:59 -> "directory-node
+    # listening" 08:57:02, so 3 seconds with 5x margin. Unlike log-node and
+    # archive-node, which were sitting at zero and one second of margin and were
+    # raised in the same change, this one has room.
+    #
+    # It is not fixed forever. Startup replays persisted claims into a fresh
+    # in-memory AKD ("replayed 20 persisted claims"), so re-measure if claim
+    # volume grows by an order of magnitude. Fly clamps this field at 60s.
     grace_period = '15s'

services/log-node/fly.toml

@@ -38,7 +38,22 @@ primary_region = 'iad'
     type = 'http'
     interval = '15s'
     timeout = '2s'
-    grace_period = '5s'
+    # grace_period MUST exceed boot-to-listening. Measured 2026-07-29 from the
+    # Fly logs: "Starting init" 08:57:02 -> "atrib-log listening" 08:57:07, so
+    # 5 seconds against a 5 second grace, i.e. no margin at all. Any slower
+    # boot fails the check on a service that is merely starting.
+    #
+    # It also grows: as the comment on [[vm]] memory below says, the persisted
+    # log is replayed into the in-memory Merkle tree at startup, so this scales
+    # with tree_size (~115k entries on the measurement date, replaying at
+    # roughly 23k entries/sec).
+    #
+    # 45s is ~9x the measured boot and covers on the order of a million entries
+    # before it needs revisiting. Fly clamps this field at 60s (`flyctl config
+    # validate` warns "this will be lowered to 1 minute"), so 60s is the hard
+    # ceiling, not a value to reach for casually: past it the only fix is
+    # making startup cheaper.
+    grace_period = '45s'
     method = 'GET'
     # /v1/pubkey always returns 200 (even with an empty tree); /v1/checkpoint
     # returns 404 until the first entry lands. With persistence in place a