@@ -0,0 +1,14 @@
+---
+'@objectstack/service-settings': patch
+---
+
+fix(service-settings): the settings audit trail records the previous value's digest in `old_hash`, and a non-secret value's digest is the declared SHA-256
+
+Clause-②: no
+Metadata-protocol: none
+
+`sys_setting_audit` and `sys_platform_setting_audit` declare `old_hash` as the SHA-256 of the previous value (null when it was unset) and `new_hash` as the SHA-256 of the new value (null on a reset). The write path passed `old_hash: null` on every write and recorded the adapter's FNV-1a 32 (`fnv32:…`) as the new hash, so an operator could not tell from the trail whether a write changed anything.
+
+- **`old_hash`** now holds the digest of the value the row held before the write, read off the row the upsert already looks up: no extra read, and none at all for a first write. It is null for a first write, after a reset, and for a row that held a secret.
+- **A non-secret value's digest** is now `sha256:` followed by the hex SHA-256 of its canonical JSON, on both `new_hash` and the `config_change` row's digest in `sys_audit_log`. It no longer goes through `CryptoAdapter.digest`, so an injected adapter cannot change the format. The prefix keeps it distinct from the `fnv32:` rows already stored, which are left as they are: compare hashes only between rows that carry the same prefix.
+- **Secret values are unchanged.** They keep the crypto provider's keyed digest (`hmac-sha256:…`) as `new_hash`, or none, and `old_hash` stays null for them.
@@ -3136,7 +3136,7 @@
       "title": "The settings audit trail keeps no fingerprint of a secret-valued setting that a reader can match offline: both ledgers record the platform's keyed digest or none, never the value or an unkeyed hash, so a caller who can only read the audit rows learns that the secret changed and nothing about its value",
       "since": "v17",
       "status": "active",
-      "revision": 4,
+      "revision": 5,
       "priority": "P1",
       "surface": "api",
       "personas": [
@@ -3159,10 +3159,10 @@
         "generate a fresh random value V1 for this run (never a value used anywhere else). As admin, PUT /api/settings/mail { provider: 'resend', api_key: V1, from_email: 'ops@example.com' }",
         "read the per-key ledger: GET /api/v1/data/sys_platform_setting_audit filtered to namespace=mail and key=api_key; capture the newest row (action set, encrypted true, new_hash). mail is a GLOBAL-scope namespace, and the per-key ledger of the global rung is the deployment-level sys_platform_setting_audit (ADR-0131 D7, packages/services/service-settings/src/settings-service-plugin.ts#settingAuditObjectFor); a tenant- or user-scope key's row lands in sys_setting_audit",
         "read the generic ledger: GET /api/v1/data/sys_audit_log filtered to action=config_change; capture the newest row for mail.api_key (object_name 'sys_platform_setting' — mail is a GLOBAL-scope namespace, and a global-scope value's row names the tenant-less table it lives in, packages/services/service-settings/src/config-change-audit.ts#CONFIG_CHANGE_GLOBAL_OBJECT_NAME; a tenant-scope key's row names 'sys_setting'; new_value is a JSON string carrying namespace, key, scope and digest; metadata carries encrypted: true)",
-        "OFFLINE, with no server key: compute the unkeyed digests a leak would carry — the FNV-1a 32 of V1's canonical JSON (the quoted string), spelled 'fnv32:' + 8 lowercase hex, which is the unkeyed digest this settings service writes for a non-secret key (packages/services/service-settings/src/crypto-adapter.ts#digest), and, belt and braces, the SHA-256 hex of V1 (bare, and with a 'sha256:' prefix) — and search both captured rows, every field, for each of them and for V1 itself",
+        "OFFLINE, with no server key: compute the unkeyed digests a leak would carry — the SHA-256 of V1's canonical JSON (the quoted string), spelled 'sha256:' + 64 lowercase hex, which is the unkeyed digest this settings service writes for a non-secret key (packages/services/service-settings/src/crypto-adapter.ts#settingAuditDigest), and, belt and braces, the bare SHA-256 hex of V1 and of its canonical JSON, and the FNV-1a 32 of its canonical JSON ('fnv32:' + 8 lowercase hex, the non-secret digest rows written before #26465 carry) — and search both captured rows, every field, for each of them and for V1 itself",
         "write V1 again, then a different value V2, then V1 a third time; re-read both ledgers and compare the fingerprints of the three writes",
         "reset the key: PUT /api/settings/mail { provider: 'log', api_key: null } (a null value is the reset; 'log' needs no api_key), then read the reset rows",
-        "control: write the NON-secret key mail.from_email and compute OFFLINE the FNV-1a 32 of its canonical JSON (the quoted string), spelled 'fnv32:' + 8 lowercase hex. Its rows must carry exactly that unkeyed digest. This proves step 4's offline search — which computes the same algorithm — WOULD have found an unkeyed digest, and that the keyed rule is scoped to secret-valued keys"
+        "control: write the NON-secret key mail.from_email and compute OFFLINE the SHA-256 of its canonical JSON (the quoted string), spelled 'sha256:' + 64 lowercase hex. Its rows must carry exactly that unkeyed digest. This proves step 4's offline search — which computes the same algorithm — WOULD have found an unkeyed digest, and that the keyed rule is scoped to secret-valued keys"
       ],
       "acceptance": [
         {
@@ -3172,7 +3172,7 @@
           "evidence": "both captured rows, quoted in full"
         },
         {
-          "clause": "a reader of the audit rows who holds no server key cannot confirm a guess: no fingerprint in either row can be reproduced from the value alone. No offline unkeyed digest of V1 — the FNV-1a 32 of its canonical JSON ('fnv32:…', the algorithm the service uses for non-secret keys), nor the SHA-256 of V1 with or without the sha256 prefix — matches anything in either row",
+          "clause": "a reader of the audit rows who holds no server key cannot confirm a guess: no fingerprint in either row can be reproduced from the value alone. No offline unkeyed digest of V1 — the SHA-256 of its canonical JSON ('sha256:…', the algorithm the service uses for non-secret keys), the bare SHA-256 of V1, nor the FNV-1a 32 of its canonical JSON ('fnv32:…') — matches anything in either row",
           "oracle": "api",
           "verify": "the step 4 offline search over both rows finds no hit, AND the step 7 control's offline digest DOES match the non-secret row. Without the control, a no-hit proves nothing about the search",
           "evidence": "the offline digest values, the two secret rows, and the matching control row"
@@ -3190,9 +3190,9 @@
           "evidence": "the two reset rows"
         },
         {
-          "clause": "non-secret settings are unchanged: a non-encrypted key keeps the unkeyed digest of its canonical JSON on both ledgers. The keyed rule applies to secret-valued keys only",
+          "clause": "non-secret settings record the unkeyed digest of their canonical JSON on both ledgers, the SHA-256 the per-key ledger's new_hash column declares. The keyed rule applies to secret-valued keys only",
           "oracle": "api",
-          "verify": "mail.from_email's rows carry the FNV-1a 32 digest of the quoted JSON string ('fnv32:' + 8 lowercase hex), computed offline in step 7",
+          "verify": "mail.from_email's rows carry the SHA-256 digest of the quoted JSON string ('sha256:' + 64 lowercase hex), computed offline in step 7",
           "evidence": "the control rows and the offline digest"
         },
         {
@@ -3203,7 +3203,7 @@
         }
       ],
       "negative": [
-        "an unkeyed digest of a secret-valued setting (the 'fnv32:' FNV-1a 32 of its canonical JSON, or a plain SHA-256 of the value, with or without a prefix) on EITHER ledger is a FAIL. Anyone who can read the audit rows could confirm a guessed value offline, and settings secrets are short, guessable values",
+        "an unkeyed digest of a secret-valued setting (the 'sha256:' SHA-256 of its canonical JSON, a plain SHA-256 of the value, or the 'fnv32:' FNV-1a 32 of its canonical JSON) on EITHER ledger is a FAIL. Anyone who can read the audit rows could confirm a guessed value offline, and settings secrets are short, guessable values",
         "the plaintext of a secret-valued setting in any field of either ledger is a FAIL",
         "a host with no keyed digest falling back to an unkeyed one is a FAIL; the rule is to record no fingerprint",
         "a secret write REFUSED because its audit fingerprint could not be computed is a FAIL: the audit trail never vetoes a settings save",
@@ -3214,7 +3214,7 @@
       ],
       "automated": {
         "kind": "unit",
-        "ref": "packages/services/service-settings/src/settings-audit-secret-digest.test.ts — acceptance[5] and negative[2]–[3] ONLY: the no-keyed-digest arm a stock boot cannot reach (knownGaps[0]), through 'legacy inline-adapter path with no provider records no fingerprint and reports it once' and 'a provider that refuses a keyed digest leaves the write intact and records no fingerprint'. The pin drives SettingsService over stub audit sinks. So its other cases — 'records the provider keyed digest on both ledgers, not the unkeyed hash of the value', 'is stable for equal values and distinct for different ones', 'records no fingerprint on a reset', the provider-wired inline-adapter case, and the non-secret 'keep the unkeyed adapter digest of the canonical JSON, unchanged' — are service-layer evidence for acceptance[0]–[4] and negative[0]–[1], never their pass: those score the ledger rows read back over the API (steps[0]–[6])"
+        "ref": "packages/services/service-settings/src/settings-audit-secret-digest.test.ts — acceptance[5] and negative[2]–[3] ONLY: the no-keyed-digest arm a stock boot cannot reach (knownGaps[0]), through 'legacy inline-adapter path with no provider records no fingerprint and reports it once' and 'a provider that refuses a keyed digest leaves the write intact and records no fingerprint'. The pin drives SettingsService over stub audit sinks. So its other cases — 'records the provider keyed digest on both ledgers, not the unkeyed hash of the value', 'is stable for equal values and distinct for different ones', 'records no fingerprint on a reset', the provider-wired inline-adapter case, and the non-secret 'record the declared SHA-256 of the canonical JSON on both ledgers, not the adapter digest (#26465)' — are service-layer evidence for acceptance[0]–[4] and negative[0]–[1], never their pass: those score the ledger rows read back over the API (steps[0]–[6])"
       },
       "source": [
         "packages/services/service-settings/src/settings-service.ts#secretAuditDigest (one helper supplies a secret write's fingerprint on both write branches, the sys_secret path and the legacy inline-adapter path: the provider's keyedDigest, or null with a once-per-key warn; never the adapter's unkeyed digest)",
@@ -3248,6 +3248,12 @@
           "date": "2026-10-20",
           "change": "automated.ref narrowed per clause — the sweep-wide rule-6 contradiction QA run #24840 (18.0 sweep round 4, Tier 1 at 307ce4c4) names at settings-audit acceptance[0]–[4], round 15 of wave anchor #24579. The ref listed seven cases whose names read as acceptance[0], [2], [3] and [4]. The pin drives SettingsService over stub audit sinks and never reads sys_platform_setting_audit or sys_audit_log over the API, which is what those clauses score. By RUNNER rule 6 (a pin scores only the layer it drives) it scores acceptance[5] and its negatives [2] and [3] only (QA run #24838 scored acceptance[0]–[4] live). Read at base b1676441ba. No clause changed",
           "ref": "#24891"
+        },
+        {
+          "revision": 5,
+          "date": "2026-10-27",
+          "change": "steps[3] and [6], acceptance[1].clause, [4] (clause, verify), negative[0] and automated.ref re-pointed: a non-secret setting's audit digest is now the SHA-256 the sys_setting_audit / sys_platform_setting_audit columns declare, spelled 'sha256:' + 64 hex (crypto-adapter.ts settingAuditDigest), no longer the adapter's FNV-1a 32 ('fnv32:'). The step-7 control computes the algorithm the service now writes, so it can still prove the offline search would find an unkeyed digest. Not loosened: the secret search keeps the FNV-1a 32 as a second net, so a secret row carrying either unkeyed form still fails",
+          "ref": "#26465"
         }
       ]
     },
 
@@ -1,5 +1,7 @@
 // Copyright (c) 2025 ObjectStack. Licensed under the Apache-2.0 license.
 
+import { createHash } from 'node:crypto';
+
 /**
  * Pluggable adapter for at-rest encryption of `Specifier.encrypted: true`
  * values. The default {@link NoopCryptoAdapter} provides a transparent
@@ -14,8 +16,14 @@ export interface CryptoAdapter {
   /** Returns the plaintext used by the resolver. */
   decrypt(ciphertext: string, ctx: { namespace: string; key: string }): Promise<string>;
   /**
-   * Stable, short, non-reversible digest used for audit-log entries so
-   * operators can correlate value changes without leaking secrets.
+   * Stable, short, non-reversible digest of a plaintext.
+   *
+   * Not read by the settings audit trail since #26465: `sys_setting_audit`
+   * and `sys_platform_setting_audit` declare SHA-256 for `old_hash` /
+   * `new_hash`, so the non-secret fingerprint is {@link settingAuditDigest}
+   * whatever adapter is injected, and a secret is fingerprinted only by the
+   * crypto provider's keyed digest. Kept on the interface so existing
+   * adapters still satisfy it.
    */
   digest(plaintext: string): string;
   /**
@@ -79,7 +87,8 @@ export class NoopCryptoAdapter implements CryptoAdapter {
     return Buffer.from(ciphertext.slice(4), 'base64').toString('utf8');
   }
   digest(plaintext: string): string {
-    // FNV-1a 32-bit — short, stable, non-cryptographic. Audit-only.
+    // FNV-1a 32-bit — short, stable, non-cryptographic. The settings audit
+    // no longer records it (#26465, see `settingAuditDigest`).
     let h = 0x811c9dc5;
     for (let i = 0; i < plaintext.length; i++) {
       h ^= plaintext.charCodeAt(i);
@@ -89,6 +98,24 @@ export class NoopCryptoAdapter implements CryptoAdapter {
   }
 }
 
+/**
+ * The fingerprint the settings audit trail records for a NON-secret value
+ * (#26465): `sha256:` followed by the hex SHA-256 of the value's canonical
+ * JSON, which is the algorithm `old_hash` / `new_hash` declare on
+ * `sys_setting_audit` and `sys_platform_setting_audit`.
+ *
+ * Computed here rather than through {@link CryptoAdapter.digest}, so an
+ * injected adapter cannot change the column's format. The prefix keeps the
+ * value self-describing beside the rows written before it (`fnv32:` from the
+ * old adapter digest) and the keyed secret fingerprints (`hmac-sha256:`), so a
+ * reader comparing two rows can tell a different algorithm from a different
+ * value. Never used for a secret: an unkeyed hash of a short secret lets any
+ * reader of the audit rows confirm a guess offline.
+ */
+export function settingAuditDigest(canonicalJson: string): string {
+  return 'sha256:' + createHash('sha256').update(canonicalJson, 'utf8').digest('hex');
+}
+
 /**
  * Is this adapter allowed to hold a declared-secret value at rest (#8026)?
  *