spec(0.10.0): §2.4 — retention_profile names behavior.ngram_size; §12.1 omits behavior when two composed inputs differ in it

CHANGELOG.md

@@ -42,7 +42,8 @@ joint probability at `ngram_size` above 2 now violates, a rule on which template
 `branching_delta` row (§13.1) that a producer emitting one for a template branching on one
 side only now violates, and a binding of each param to a whole-token wildcard (§3.5) that a
 producer emitting a param for a wildcard inside a token, or none for a whole-token wildcard,
-now violates.
+now violates, and two rules on `behavior.ngram_size` (§2.4, §12.1) that a producer keeping one
+`retention_profile` across two orders, or a composer merging two orders' blocks, now violates.
 It also adds one optional member
 (`withheld_signals`, §13.2.2), which is additive on its own. MINOR bump: MAJOR stays `0`, so
 §6's *"the MAJOR field of `metalog_version` must equal the MAJOR of the spec"* is
@@ -71,10 +72,36 @@ in pull request [#14](https://github.com/CodeRoasted/metalog-spec/pull/14) under
 0.x line without an `rfc:` issue or a comment window, while the reference
 implementation is the only producer and consumer. The RFC comes back at the v1.0
 freeze, or earlier the day a second implementation is listed. P1, P3 and P4 stay
-proposed; any of them that is taken lands as an editor change recorded here.
+proposed; any of them that is taken lands as an editor change recorded here. **P3's item (f)**,
+the order-dependent `behavior.ngram_size`, is the first taken: see the §2.4 / §12.1 entry under
+*Changed*. The rest of P3 (the caps rule) stays proposed.
 
 ### Changed
 
+- **§2.4 — `retention_profile` names `behavior.ngram_size`, and §12.1 omits `behavior` when
+  two composed inputs differ in it.** `ngram_size` was in neither identifier's list, so two
+  documents produced at different sequence orders could carry equal identifiers and pass
+  §2.4's comparability gate. `compose()` then had to pick one input's `ngram_size` — a
+  required field — so `compose(A, B)` and `compose(B, A)` declared different values, against
+  §12.2's commutativity MUST; and a `MetaLogDiff` across the pair reported every n-gram of
+  one side as vanished and every n-gram of the other as new, which §13.2 counts as a witness
+  of `"changed"`. An order-`m` key and an order-`n` key denote different objects, so no
+  minimum repairs it: the smaller order either declares an order the array does not hold or
+  drops one input's n-grams with no field reporting the loss. Two changes. **§2.4** states
+  what `retention_profile` names as a rule — every parameter that fixes which entries a
+  bounded block retains, how they are ranked, and what an entry's key denotes — with the
+  existing items and `behavior.ngram_size` as its examples, so the existing gate refuses a
+  cross-order `compose()` and `MetaLogDiff` at both sites with no new mechanism. **§12.1**
+  covers the case §2.4's gate does not reach (an input without the identifier): when both
+  inputs carry `behavior` at different `ngram_size`, `C.behavior` **MUST** be omitted, never
+  merged at a minimum and never carried from one side; when they agree, `C.behavior.ngram_size`
+  is that value. Omitting the block also drops `dropped_ngram_observations`: a count lost,
+  not falsified. The §1 table and the schema's `retention_profile` description name the
+  parameter. **Breaking** — a producer keeping one `retention_profile` across two orders, or a
+  composer merging two orders' blocks, becomes non-conformant — and editor-merged under
+  [`GOVERNANCE.md`](GOVERNANCE.md) §2's 0.x rule. No document's validity changes: both
+  identifiers stay opaque strings and `behavior` stays optional. RFC #8, P3 (f).
+
 - **A window with no event time has no envelope (§2.2).** `window.start`, `window.end` and
   `window.duration_seconds` are present together, or absent together when the producer
   attributes an event time to none of the window's lines — a window with no line

SPEC.md

@@ -78,7 +78,7 @@ top-level fields:
 | `window` | object | yes | The line count and, when any line has one, the event-time envelope of the lines the window contains. See §2.2. |
 | `source` | object | yes | What was observed (service, host, fleet). See §2.3. |
 | `canonicalization_version` | string | no | Opaque identifier for the canonicalization rules in effect. Gates `compose()`/diff comparability. See §2.4. |
-| `retention_profile` | string | no | Opaque identifier for the retention parameters (top_k, reservoir, salience weights, diversity caps). Gates `compose()`/diff comparability. See §2.4. |
+| `retention_profile` | string | no | Opaque identifier for the retention parameters (top_k, reservoir, salience weights, diversity caps, `behavior.ngram_size`). Gates `compose()`/diff comparability. See §2.4. |
 | `stats` | object | yes | Per-template counts and frequency metrics. See §3. |
 | `templates` | object | no | Optional dedup map `template_id → template_str`. See §3.4. **RESERVED** — see §2. |
 | `behavior` | object | no | Sequence/transition fingerprint. See §4. |
@@ -233,10 +233,14 @@ A MetaLog **MAY** carry two opaque processing-identifier strings that name the
   templates and structural metadata). It **MUST** be bumped when those rules'
   *output-affecting* semantics change; a binary rebuild with no rule change
   **MUST NOT** bump it. It is **not** a binary build id.
-- `retention_profile` — names the **retention parameters** in effect: `top_k`
-  size (§3.1), reservoir admission weights and size and diversity caps (§3.7),
-  and the salience arithmetic. It **MUST** be bumped when any of those
-  parameters change.
+- `retention_profile` — names the **retention parameters** in effect: every
+  parameter that fixes which entries a bounded block retains, how they are ranked,
+  and **what an entry's key denotes**. At this version those are `top_k` size
+  (§3.1), the reservoir admission weights, size and diversity caps and the salience
+  arithmetic (§3.7), and `behavior.ngram_size` (§4) — the sequence order, which
+  fixes the *domain* of `top_ngrams` rather than a bound on it. That list
+  illustrates the rule and does not close it: a parameter meeting the rule joins it
+  without an edit here. It **MUST** be bumped when any such parameter changes.
 
 The values are **opaque strings**. This spec defines neither a registry of names
 nor a canonical format; producers and consumers within an environment **MUST**
@@ -1364,6 +1368,21 @@ to a 1-hour MetaLog).
 - `C.behavior.dominant_path` is re-derived greedily from the merged
   graph; consumers **MUST NOT** assume it equals the path of either
   input.
+- `C.behavior` **MUST** be omitted entirely when both inputs carry a `behavior` block
+  and their `ngram_size` values **differ**. An order-`m` key and an order-`n` key with
+  `m ≠ n` denote different objects, so there is no merged n-gram population to derive:
+  counts across the two orders are not comparable (every order-`n` occurrence implies
+  an order-`m` one for `m < n`, so a joint ranking is biased toward the shorter order)
+  and `probability` is conditioned to a different depth on each side. Where the two
+  agree, `C.behavior.ngram_size` is that value. A composer **MUST NOT** instead take
+  the smaller order, nor carry one input's block through unchanged: the first declares
+  an order the merged array does not hold or silently drops one input's n-grams, and
+  the second describes one input's lines under a `window` covering both. Omitting the
+  block also drops `dropped_ngram_observations`; that **loses** a count and does not
+  falsify one, because §4's omission rule governs the field inside a present block,
+  not the block. The precondition is block-grained, like the `cube` clause below: the
+  rest of `C` composes as usual. It is the residual case only — whenever both inputs
+  carry a `retention_profile`, §2.4's gate refuses the pair before composition begins.
 - `C.stability` **MUST** be omitted (it is meaningless across
   composed inputs); consumers wanting a current-vs-prior view of a
   composed document should use §13 Diff explicitly.

schema/metalog.v0.schema.json

@@ -467,7 +467,7 @@
         },
         "retention_profile": {
             "type": "string",
-            "description": "Opaque identifier for the retention parameters — top_k size, reservoir admission weights/size/diversity caps, and the salience arithmetic; gates compose()/diff comparability (SPEC §2.4). Added in v0.5.0."
+            "description": "Opaque identifier for the retention parameters — top_k size, reservoir admission weights/size/diversity caps, the salience arithmetic, and behavior.ngram_size; gates compose()/diff comparability (SPEC §2.4). Added in v0.5.0."
         },
         "coordinate": {
             "description": "Re-derivation coordinate addressing this window back to its source: raw(window) = replay(source, bounds) (SPEC §15). A raw coordinate XOR a composed coordinate. Added in v0.5.0.",