efa: trial the leverage pass on three sections, with before and after scores

First run of the leverage pass, measured rather than assumed. Three sections
scored by a skeptical staff-engineer reader before and after, on mental model,
practical leverage, rigor, accuracy and economy. The after-scoring was done by a
fresh reader that never saw the baseline, to avoid anchoring.

                model  leverage  rigor  accuracy  economy   total
  EFA Device     3->5     5->5    4->4     5->5     3->4   20->23
  libfabric      4->4     4->5    4->4     5->5     4->3   21->21
  SRD            5->5     3->4    5->3     5->5     3->3   21->20

The result is not a uniform win, and that is the useful part.

EFA Device gained three points. It was already forward-oriented, so the pass had
something to build on: the four counting layers became the spine and the model
is now demonstrated predictive on an instance type the page does not start from,
checked by 16 x 400 = 6,400 Gbps. It also lost a provenance defect, where an
inference carried a code badge pointing at a rail-sorting comment that said
nothing about the claim.

libfabric was a wash. It bought a point of leverage and spent a point of economy,
and grew 61 lines against an instruction to hold or fall. The protocol ladder is
now actionable, which was the goal, but the section still delivers its payload at
the end of every container.

SRD went backwards. It gained leverage, from a five-code completion-error triage
table built from newly fetched provider sources, and lost two points of rigor
because the pass introduced two universal negatives cited to functions that
cannot evidence them: "no retransmission logic exists anywhere in the EFA driver"
attached to a stats-copying function, and "no matching set function anywhere in
the tree" attached to the getter. Both are now rescoped to what the search
actually established. The standard says the pass may trade rigor for reach by
dropping citations on undisputed claims. It does not license new claims that
outrun their evidence, and the wording needs to say so.

Accuracy held at five everywhere, which was the hard constraint.

Also fixed, from the same review: the driver SHA was described three different
ways across the dive. Checked against the API, the tag efa_linux_3.3.0 points at
b99452b7, no bare r3.3.0 tag exists, and master has moved past it. All three
descriptions now name the real tag.

Not rolled out to the other seventeen sections. On this evidence the pass helps
most where a section is already forward-oriented, and is riskiest on the most
rigorous material, which is the opposite of where it looked most needed.

Gates: pnpm gates exit 0, Playwright 8 of 8 green, verify-citations exit 0 with
157 URLs and 170 pinned files resolving, 12 tests passing.

Author: Carlos Manzanedo Rueda

deep-dives/efa/revamp/leverage-after-2026-08-05.md

@@ -0,0 +1,94 @@
+# The Leverage Pass
+
+Established 2026-08-05 by Carlos. The last pass over a dive, run after research,
+verification and editorial shaping are done.
+
+## Why it exists
+
+Rigor is what makes the research trustworthy. It is not what makes the dive
+worth reading. This project traded the outcome for the rigor: a meticulously
+sourced artifact that a reader admires and cannot act on. This pass buys the
+outcome back, and it spends some rigor to do it.
+
+The rigor stays in the research. It does not all need to reach the page.
+
+## The two modes, and why oscillating between them is correct
+
+Deep dives get written in two directions, and both are legitimate:
+
+**Backward.** Here is a technology. How does it work, where does it fit, what
+would I use it for. This is how the research runs, and it produces mechanism.
+
+**Forward.** Here is what I am trying to do. What do I need, and what decisions
+does it force. This is how a reader arrives, and it produces consequence.
+
+The research must run backward, because you cannot reason about a system you
+have not taken apart. The page must read forward, because that is the direction
+the reader is travelling. Most of the weakness in a finished section is
+backward-mode material left in backward-mode order.
+
+Corollary for how standards get written: defining what you do NOT want is the
+right tool while exploring, because it preserves the surface you have not
+decided about yet. Defining what you DO want is the right tool at consolidation,
+because it forces the outcome. Using the positive form too early narrows the
+search; using the negative form too late leaves the reader to infer the point.
+
+## The pass
+
+For every substantial block, answer three questions in order. Blocks that
+answer none come out, however hard they were to establish.
+
+**1. What does this mean to the reader and to the document?**
+Say why the block is here, in one sentence, in terms of the reader rather than
+the subject. A block that can only be justified by "it is true and it was hard
+to find" has failed.
+
+**2. What mental model does it build, and is that model visible?**
+Name the model in one line. Then ask whether a diagram or a visualization would
+carry it better than the prose does. A mental model that only exists as
+paragraphs is usually a mental model the reader will not retain. This is where
+new diagrams get commissioned, and it is a legitimate reason to add rather than
+cut.
+
+**3. How would the reader leverage this?**
+The hardest question and the one most often skipped. Concretely: what does it
+let them decide, configure, size, choose between, or monitor. Monitoring is the
+underrated answer. A mechanism that explains a counter, a log line, a failure
+signature or a knob has an obvious practical hook, and mechanism sections
+frequently have one that was never stated.
+
+If a block builds a strong model but has no leverage, keep it and say plainly
+what it is for: understanding the system, so that the practical sections make
+sense. That is an honest and sufficient answer. What is not sufficient is
+leaving the reader to work out which kind of block they are reading.
+
+## What this pass is allowed to trade
+
+Rigor for reach, deliberately and visibly:
+
+- Citations on claims nobody disputes can go. The claim stays.
+- Evidence that exists to prove we did the work, rather than to help the reader
+  check ours, can go behind a disclosure or out entirely.
+- Depth established during research that answers no reader question comes out,
+  and the research file keeps it.
+
+What this pass never trades: a sourced number, a documentation-versus-code
+conflict, a hedge scope, or an UNKNOWN.
+
+## How it is measured
+
+The pass is evaluated, not assumed. Before and after, each section is scored as
+a reader would, one to five on each axis:
+
+- **Mental model.** Does it give a way to think about the system and predict
+  what it will do?
+- **Practical leverage.** Does it change what the reader decides, configures,
+  sizes or monitors?
+- **Rigor.** Is every claim sourced, scoped and correctly attributed?
+- **Accuracy.** Is it right?
+- **Economy.** Is the value dense, or diluted by material serving neither model
+  nor practice?
+
+A successful pass raises model and leverage, holds accuracy at five, holds or
+improves economy, and may spend a point of rigor. Losing accuracy fails the
+pass regardless of the other scores.