docs: re-measure the corpus-access table against the published corpus (#56)

#56 step 4 asked for the README table to be re-measured after the pass. The old
figures predate the row-group rechunk (cog_pipeline#93, published 2026-08-09)
and reported the mirrored column as 'local speed' with no number -- hiding the
largest difference available to a user.

Measured 2026-08-10, fresh R session per arm, against the live corpus at
pipeline_commit 3d28ddd. Madison WI, 16-core Linux workstation.

Three findings the old table could not express:

- A local mirror is 60-80x faster. A one-off question is ~12 s end to end
  remotely against ~0.15 s mirrored. Stated outright now, because it is a
  bigger and cheaper win for users than anything in the R code.

- Opening the session is the LARGEST remote cost (~7.5 s), bigger than any
  individual query, and it lands on the user's first query rather than on
  library(). The old table accounted for it nowhere, so every per-query figure
  was quietly missing it.

- The remote cost is round-trips, not scanning: a repeat query over
  already-touched partitions is ~1.5 s against ~4 s cold, and a full-history
  query costs ~7 s whether it runs first or last (verified by running the arms
  in both orders). This is why #93's 1.4-1.7x, measured through cog-api against
  a local mount, does not show up on the remote path -- there, network latency
  swamps scan time.

Corpus size corrected to ~201 MB: row-group chunking added ~3.4%, and 190.6 was
ambiguous between MB and MiB besides. Measured from the manifest and on disk.
The 0.3.0 NEWS section keeps 190.6 -- it was correct for that release.

Also documented HTTP 429: a burst of remote queries gets rate-limited by the
host. Hit while taking these measurements.

NEWS.md

@@ -1,5 +1,29 @@
 # uscogdata 0.4.0
 
+## Documentation: the corpus-access table is re-measured and honest
+
+The README's "two ways to read the corpus" table carried figures taken before
+the corpus was re-chunked into row groups (cog_pipeline#93, published
+2026-08-09) and reported the mirrored column as "local speed" with no number at
+all. Re-measured 2026-08-10 against the published corpus (`pipeline_commit
+3d28ddd`), fresh R session per arm:
+
+* **A local mirror is roughly 60-80x faster.** A one-off question costs ~12 s
+  end to end remotely against ~0.15 s mirrored. That is the largest single
+  difference available to a user and it is now stated outright rather than left
+  as "local speed".
+* **Opening the session is the largest remote cost** (~7.5 s -- manifest fetch
+  plus 23 view registrations over HTTPS), larger than any individual query, and
+  it lands on the first query rather than on `library(uscogdata)`. The old table
+  did not account for it anywhere.
+* **The remote cost is round-trips, not scanning.** A repeat query over
+  already-touched partitions is ~1.5 s against ~4 s cold, and a full-history
+  query costs ~7 s whether it runs first or last.
+* The corpus size is **~201 MB**, not 190.6 MB -- row-group chunking added ~3.4%
+  and the old figure was ambiguous between MB and MiB besides.
+* Documented that a burst of remote queries can be rate-limited by the host
+  (`HTTP 429`), which is another reason to mirror for real work.
+
 ## Cohorts can be named by predicate, not just by id
 
 `cog_spending()`, `cog_revenue()` and `cog_balances()` gain optional `state`

README.md

@@ -18,7 +18,7 @@ carries provenance describing what was converted, what was aggregated, and
 which known series breaks intersect your query.
 
 **Scope:** government types 0–3 (state, county, municipality, township).
-56 fiscal years, 46,148,034 rows, 190.6 MB. There is no source data for FY1968
+56 fiscal years, 46,148,034 rows, ~201 MB. There is no source data for FY1968
 or FY1969. Special districts (type 4) and school districts (type 5) are
 excluded pending validation.
 
@@ -85,14 +85,38 @@ cog_explain(spend)
 
 | | Remote (default) | Mirrored |
 |---|---|---|
-| Setup | none | `cog_mirror(dest)`, 190.6 MB once |
-| Disk used | **0 MB** — HTTP range requests only | 190.6 MB |
-| Per query | ~4 s (one government, one year)<br>~6 s (one government, 23 years) | local speed |
+| Setup | none | `cog_mirror(dest)`, ~201 MB once |
+| Disk used | **0 MB** — HTTP range requests only | ~201 MB |
+| Opening a session | ~7.5 s | ~0.1 s |
+| One government, one year | ~4 s | ~0.05 s |
+| One government, full history | ~7 s | ~0.1 s |
+| Later queries, same session | ~1.5 s | ~0.05 s |
 | Good for | trying it out, teaching, one-off questions | repeated analysis, offline work, reproducibility |
 
+**A local mirror is roughly 60–80x faster, and it is one function call.** That is
+by far the largest difference any of these settings makes. If you are going to
+ask more than a handful of questions, mirror first.
+
+Measured 2026-08-10 on a 16-core Linux workstation against the published corpus
+(schema v7, `pipeline_commit 3d28ddd`), fresh R session per arm. A one-off
+question costs about **12 seconds end to end remotely and 0.15 seconds
+mirrored**, session setup included.
+
+Two things the per-query rows hide:
+
+- **Opening the session is the single largest remote cost** — larger than any
+  one query. It fetches the manifest and registers 23 SQL views over HTTPS, and
+  it lands on your first query, not on `library(uscogdata)`.
+- **The cost is network round-trips, not scanning.** A repeat query against
+  partitions this session has already touched is ~1.5 s rather than ~4 s, and a
+  full-history query costs ~7 s whether it runs first or last. What you are
+  paying for is reaching each of the 56 yearly files over HTTPS the first time.
+
 Nothing is written to disk in remote mode: DuckDB fetches the parquet footer,
 works out which row groups it needs, and reads only those. Nothing is cached
-between sessions either, so every query goes back to the network.
+between sessions either, so every query goes back to the network — and a session
+that issues many remote queries in quick succession can be rate-limited by the
+host (`HTTP Error: ... 429`). Both are further reasons to mirror for real work.
 
 The default points at a public HuggingFace mirror of the corpus. If you would
 rather not depend on a third party — for reproducibility, for an air-gapped