diff --git a/AGENTS.md b/AGENTS.md
index dc44c8d..6f4f521 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -21,6 +21,7 @@ This is the contributor guide for AI agents (and humans!) modifying this repo. T
 - `nsc/cache/` — disk cache for generated command-models.
 - `nsc/auth/` — login verification helpers (pre-flight probes, token rotate).
 - `nsc/aliases/` — curated alias resolver (`ls`, `get`, `rm`, `search`). Framework-free.
+- `nsc/completion/` — dynamic shell-completion: framework-free providers + cheap on-disk cache probe + Typer `shell_complete` callbacks.
 - `nsc/skill/` — bundle-path helper for the portable `SKILL.md` shipped in the wheel.
 - `nsc/schemas/bundled/` — versioned NetBox OpenAPI snapshots, fallback when offline.
 
diff --git a/CHANGELOG.md b/CHANGELOG.md
index c8d82df..22d427a 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,63 @@
 
 All notable changes to netbox-super-cli are tracked here. Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/) loosely. From v1.0.0 onward, releases follow [Semantic Versioning](https://semver.org/) and the version in `pyproject.toml` matches the git tag. Pre-1.0 milestones (Phase 1-5) were pinned by tag while `pyproject.toml` stayed at `0.0.1`.
 
+## v1.3.0 — unreleased
+
+Minor release. A batch of user-facing features — singular alias forms, parallel
+bulk writes, dynamic shell completion, and a strict `audit_redaction: full`
+mode — plus a round of state-directory, workflow, and supply-chain hardening.
+
+### Added
+
+- **Singular alias forms** ([#6]). `nsc ls device` now resolves like
+  `nsc ls devices` for a curated set of nine resources: `device`, `prefix`,
+  `tenant`, `vlan`, `site`, `rack`, `interface`, `cable`, and `tag`. A
+  non-curated singular that does not resolve suggests the plural
+  (e.g. "Did you mean `devices`?") rather than guessing.
+- **`--workers N` on bulk writes** ([#3]). Bulk write commands
+  (`create --ndjson`, `update --ndjson`, …) accept `--workers N` to run up to N
+  concurrent in-flight requests (default `1`, max `32`). Per-record
+  `--on-error` semantics are preserved, and the audit log stays one well-formed
+  line per record (writes are thread-safe). With `--on-error stop`, the bound
+  applies to *new* submissions — in-flight requests (up to N) are allowed to
+  complete.
+- **Dynamic shell completion** ([#2]). Completion now reads the cached schema to
+  complete dynamic values: `nsc ls dev<TAB>` →
+  `devices device-roles device-types`; `nsc --profile <TAB>` → profile names;
+  `nsc <tag> <resource> list --status <TAB>` → enum values. This complements the
+  existing static `--install-completion` stubs.
+- **`audit_redaction: full` config mode** ([#5]). A new strict redaction mode
+  redacts ALL request/response bodies, leaving only
+  `{method, url, status_code, timestamp, profile}` per audit line. The `url` is
+  sanitized — the query string and any `user:pass@` credentials are stripped.
+  The default stays `safe`.
+
+### Changed
+
+- **`~/.nsc` state root hardened to `0700`** ([#90]). Defense-in-depth on the
+  state directory itself; config and audit files were already `0600` and
+  `logs/` already `0700`.
+- **CI runners on Node.js 24** ([#9]). Bumped `actions/checkout` to v5,
+  `astral-sh/setup-uv` to v7, and `deploy-pages` to v5.
+- **Least-privilege workflow permissions** ([#12]). Every GitHub Actions
+  workflow now declares an explicit least-privilege `permissions:` block.
+- **Release workflow actions pinned to commit SHAs** ([#16]). `release.yml`
+  pins `checkout` and `setup-uv` to commit SHAs (the PyPA publish action stays
+  on its floating tag per PyPA guidance); added `.github/dependabot.yml`
+  (github-actions ecosystem, weekly).
+- **E2E CI matrix over NetBox v4.5 and v4.6** ([#4]). The end-to-end job now
+  runs against both NetBox lines (`fail-fast: false`, per-version labels).
+
+[#6]: https://github.com/thomaschristory/netbox-super-cli/issues/6
+[#3]: https://github.com/thomaschristory/netbox-super-cli/issues/3
+[#2]: https://github.com/thomaschristory/netbox-super-cli/issues/2
+[#5]: https://github.com/thomaschristory/netbox-super-cli/issues/5
+[#90]: https://github.com/thomaschristory/netbox-super-cli/issues/90
+[#9]: https://github.com/thomaschristory/netbox-super-cli/issues/9
+[#12]: https://github.com/thomaschristory/netbox-super-cli/issues/12
+[#16]: https://github.com/thomaschristory/netbox-super-cli/issues/16
+[#4]: https://github.com/thomaschristory/netbox-super-cli/issues/4
+
 ## v1.2.1 — 2026-06-25
 
 Patch release. Fixes foreign-key handling in the interactive TUI.
diff --git a/CLAUDE.md b/CLAUDE.md
index 087561a..3705dbe 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -14,6 +14,7 @@ This is the contributor guide for AI agents (and humans!) modifying this repo. T
 - `nsc/cache/` — disk cache for generated command-models.
 - `nsc/auth/` — login verification helpers (pre-flight probes, token rotate).
 - `nsc/aliases/` — curated alias resolver (`ls`, `get`, `rm`, `search`). Framework-free.
+- `nsc/completion/` — dynamic shell-completion: framework-free providers + cheap on-disk cache probe + Typer `shell_complete` callbacks.
 - `nsc/skill/` — bundle-path helper for the portable `SKILL.md` shipped in the wheel.
 - `nsc/schemas/bundled/` — versioned NetBox OpenAPI snapshots, fallback when offline.
 
diff --git a/README.md b/README.md
index 289c2fe..3212c90 100644
--- a/README.md
+++ b/README.md
@@ -74,6 +74,10 @@ uv run nsc dcim devices create -f device.yaml --apply
 uv run nsc dcim devices create -f devices.yaml --apply
 uv run nsc dcim devices create -f devices.yaml --no-bulk --on-error continue --apply
 
+# Speed up a loop-mode bulk write with concurrent requests (default 1, max 32).
+# Per-record --on-error semantics hold; the audit log stays one line per record.
+uv run nsc dcim devices create -f devices.ndjson --workers 8 --apply
+
 # NDJSON / JSONL — one record per line; parse failures abort the whole batch
 # before any wire request fires (`type: input_error`, exit 4, `details.bad_lines`).
 uv run nsc dcim devices create -f devices.ndjson --apply
@@ -108,7 +112,26 @@ Stdin is sniffed from the first 512 bytes (first non-whitespace byte plus a one-
 - HTTP headers in the `SENSITIVE_HEADERS` set (e.g. `Authorization`, `X-API-Key`) — replaced with `"<redacted>"`.
 - Request-body fields whose OpenAPI definition has `format: password` OR whose name (case-insensitive) is one of: `password`, `secret`, `token`, `api_key`, `apikey`, `private_key`, `passphrase`, `client_secret`. Nested fields and arrays of objects are walked recursively.
 
-The wire body sent to NetBox is **not** redacted — only the audit log. A failed write still records the redacted body; redaction is irreversible. Treat `audit.jsonl` like a verbose application log: gate it behind your home-directory permissions and rotate / archive accordingly. A "redact everything" mode is on the post-v1.0 roadmap.
+The wire body sent to NetBox is **not** redacted — only the audit log. A failed write still records the redacted body; redaction is irreversible. Treat `audit.jsonl` like a verbose application log: gate it behind your home-directory permissions and rotate / archive accordingly.
+
+For a stricter posture, set `audit_redaction: full` in `~/.nsc/config.yaml` (the default is `safe`, the field-level redaction described above). In `full` mode every audit line keeps only `{method, url, status_code, timestamp, profile}` — all request and response bodies are dropped, and the `url` is sanitized (query string and any `user:pass@` credentials stripped).
+
+## Curated aliases
+
+`nsc ls`, `nsc get`, `nsc rm`, and `nsc search` are short verbs over the dynamic
+command tree.
+
+```sh
+nsc ls devices                 # list a resource by plural name
+nsc ls device                  # singular also works (curated set, see below)
+nsc get devices 7              # one record by id or name
+nsc rm devices 7 --apply       # delete (dry-run by default)
+nsc search "rack-42"           # /api/search/ on NetBox 4.5+
+```
+
+Singular forms resolve for a curated set of nine resources: `device`, `prefix`,
+`tenant`, `vlan`, `site`, `rack`, `interface`, `cable`, and `tag`. Any other
+singular that does not resolve suggests its plural ("Did you mean `devices`?").
 
 ## Output and errors
 
@@ -154,7 +177,16 @@ nsc --install-completion         # auto-detects $SHELL
 nsc --show-completion            # prints the script instead of installing
 ```
 
-Typer supports `bash`, `zsh`, `fish`, and `pwsh`. Completion of dynamic values (resource names, profile names, filter keys) is on the post-1.0 roadmap.
+Typer supports `bash`, `zsh`, `fish`, and `pwsh`.
+
+Completion of dynamic values reads the cached schema and complements the static
+stubs above:
+
+```sh
+nsc ls dev<TAB>                               # devices device-roles device-types
+nsc --profile <TAB>                           # configured profile names
+nsc dcim devices list --status <TAB>          # enum values for the field
+```
 
 ## Bundled Skill for AI agents
 
diff --git a/docs/architecture/caching.md b/docs/architecture/caching.md
index 18d3a40..ee34375 100644
--- a/docs/architecture/caching.md
+++ b/docs/architecture/caching.md
@@ -70,3 +70,12 @@ nsc cache prune --apply              # actually delete
 nsc cache prune --max-age 30 --apply
 nsc cache prune --output json        # structured envelope
 ```
+
+## Shell completion reads the cache
+
+Dynamic shell completion (`nsc/completion/`) is backed by this same on-disk
+cache. At TAB time the completion providers probe the cached command-model for
+the active profile (`<profile>/<schema-hash>.json`) — never the network — so
+resource-name, profile, and enum suggestions are cheap. A missing or corrupt
+cache degrades to no suggestions rather than blocking the prompt; the cache is
+populated by the first real command (or `nsc login --fetch-schema`).
diff --git a/docs/architecture/http-client.md b/docs/architecture/http-client.md
index 17b4e1c..57107ed 100644
--- a/docs/architecture/http-client.md
+++ b/docs/architecture/http-client.md
@@ -22,6 +22,17 @@
 - Pagination helper that follows `next` URLs (used by `--all`).
 - Audit log appender at `~/.nsc/logs/audit.jsonl` — written for writes
   always, and for any request when `--debug` is set.
+
+## Concurrency (`--workers N`)
+
+Bulk write commands accept `--workers N` (default 1, max 32) to keep up to N
+requests in flight. Concurrency is **thread-based**: a `ThreadPoolExecutor`
+fans the per-record loop out over the single sync `httpx.Client` — there is no
+async path. Per-record `--on-error` semantics are preserved regardless of
+worker count. Audit appends are serialized by a module-level lock
+(`_APPEND_LOCK` in `nsc/http/audit.py`) wrapped around the whole
+open/write/close, so concurrent workers can never interleave a partial line —
+each record is one well-formed JSON line.
 - A "last request" snapshot at `~/.nsc/logs/last-request.json`
   (overwritten every call, regardless of `--debug`) — handy for triage.
 
@@ -55,9 +66,26 @@ Each line of `audit.jsonl` is a JSON object with:
 - `duration_ms`, `attempt_n`, `final_attempt`, `error_kind`
 - `dry_run`, `preflight_blocked`, `record_indices`, `applied`, `explain`
 
+## Redaction modes
+
+Redaction is applied when each entry is serialized (`nsc/http/audit.py`), so a
+failed retry never unredacts. The `defaults.audit_redaction` config setting
+selects the mode:
+
+- `safe` (default) — the full audit shape above, with sensitive headers and
+  `sensitive_paths` body fields masked to `"<redacted>"` and bodies over 256 KB
+  truncated.
+- `full` — compliance escalation that drops every body, header, and query
+  string entirely. Each line keeps exactly five keys —
+  `{method, url, status_code, timestamp, profile}` — and `url` is sanitized to
+  scheme + host[:port] + path so neither query params nor `user:pass@` userinfo
+  can leak through the one remaining string field.
+
 The audit file is append-only and rotates to `audit.jsonl.1` at 10 MB; it is
 created owner-only (`0600`) inside a `0700` logs directory, and failed writes do
-NOT unredact. See
+NOT unredact. The state root (`~/.nsc`) and its subdirectories are clamped to
+`0700` via a shared `ensure_private_dir()` (`nsc/config/settings.py`) used by
+both the config writer and the audit-dir code. See
 [Writes and safety](../guides/writes-and-safety.md) for the full redaction
 contract.
 
diff --git a/docs/architecture/overview.md b/docs/architecture/overview.md
index 629f2d3..804ef9f 100644
--- a/docs/architecture/overview.md
+++ b/docs/architecture/overview.md
@@ -17,6 +17,7 @@ nsc/
 ├── cache/           # On-disk cache for generated CommandModels
 ├── auth/            # Login verification helpers (verify probe, token rotate)
 ├── aliases/         # Curated alias resolver (ls/get/rm/search)
+├── completion/      # Dynamic shell-completion: providers + cache probe + Typer callbacks
 ├── skill/           # Bundle-path helper for the portable SKILL.md
 └── schemas/bundled/ # Versioned NetBox OpenAPI snapshots (offline fallback)
 ```
diff --git a/docs/contributing/development.md b/docs/contributing/development.md
index ed3d7ad..d51f234 100644
--- a/docs/contributing/development.md
+++ b/docs/contributing/development.md
@@ -41,6 +41,23 @@ just docs-build    # build the site, fail on broken links / missing nav
 underlying issue and create a new commit. If `ruff format` modifies a file
 during the hook, re-`git add` and re-commit.
 
+## CI
+
+GitHub Actions mirrors the local `just` targets, plus an end-to-end suite:
+
+- **E2E is multi-version.** The e2e workflow runs a matrix over NetBox **4.5
+  and 4.6** (latest patch of each minor) with `fail-fast: false`, so a
+  regression on one version surfaces as a per-dimension failure rather than a
+  sweeping red.
+- **Runners are on Node.js 24** — the JS-based actions are pinned to their
+  Node-24 majors (`actions/checkout@v5`, `astral-sh/setup-uv@v7`,
+  `actions/deploy-pages@v5`).
+- **Least-privilege permissions.** Every workflow declares an explicit
+  top-level `permissions:` block; the release workflow additionally SHA-pins
+  its actions (the PyPA publish action excepted).
+- **Dependabot** watches the `github-actions` ecosystem on a weekly schedule
+  (`.github/dependabot.yml`) so action pins stay current.
+
 ## Where the design lives
 
 - `docs/superpowers/specs/` — design specs (gitignored; local-only).
diff --git a/docs/getting-started/concepts.md b/docs/getting-started/concepts.md
index 33c0b03..3d87616 100644
--- a/docs/getting-started/concepts.md
+++ b/docs/getting-started/concepts.md
@@ -49,6 +49,12 @@ endpoints — without per-plugin code in `nsc`**.
 
 A short curated alias layer sits on top: `nsc ls`, `nsc get`, `nsc rm`, `nsc search`.
 
+For the curated resources — `device`, `prefix`, `tenant`, `vlan`, `site`,
+`rack`, `interface`, `cable`, `tag` — the singular works identically to the
+plural: `nsc ls device` is the same as `nsc ls devices` (likewise for `get` and
+`rm` where supported). A non-curated singular that doesn't resolve suggests the
+plural instead ("Did you mean `devices`?").
+
 ## Dry-run / apply
 
 Writes default to dry-run. The CLI shows the resolved request and exits 0
diff --git a/docs/getting-started/first-run.md b/docs/getting-started/first-run.md
index b378381..9432de6 100644
--- a/docs/getting-started/first-run.md
+++ b/docs/getting-started/first-run.md
@@ -74,6 +74,7 @@ nsc dcim devices list                       # default page (page_size=50)
 nsc dcim devices list --all                 # paginate to completion
 nsc dcim devices get 7                      # get by id
 nsc ls devices                              # alias — matches devices.list
+nsc ls device                               # singular works too for curated resources
 nsc dcim devices list --output json --all   # canonical machine-readable form
 ```
 
diff --git a/docs/getting-started/install.md b/docs/getting-started/install.md
index c8bab99..19ee45e 100644
--- a/docs/getting-started/install.md
+++ b/docs/getting-started/install.md
@@ -50,8 +50,16 @@ nsc --install-completion=pwsh
 
 Restart your shell, then verify with `nsc <TAB><TAB>`.
 
-> Dynamic completion (resource names, profile names, filter keys) is on the
-> post-1.0 roadmap.
+### Dynamic completion
+
+Once completion is installed, TAB expands real values from the cached schema —
+no network call at TAB time:
+
+```sh
+nsc ls dev<TAB>                  # → devices device-roles device-types
+nsc --profile <TAB>             # → profile names from ~/.nsc/config.yaml
+nsc dcim devices list --status <TAB>   # → schema enum values for that filter
+```
 
 ## Requirements
 
diff --git a/docs/guides/ci-and-automation.md b/docs/guides/ci-and-automation.md
index 3d9f452..bc4bc27 100644
--- a/docs/guides/ci-and-automation.md
+++ b/docs/guides/ci-and-automation.md
@@ -21,6 +21,22 @@ This:
 - Emits machine-readable output to stdout.
 - Returns a stable exit code.
 
+## Faster bulk loads with `--workers`
+
+When a payload loops record-by-record, `--workers N` runs up to N requests
+concurrently (default `1`, max `32`), which can sharply cut wall-clock time for
+large NDJSON loads:
+
+```sh
+nsc dcim devices create -f new-devices.ndjson --apply --workers 8 \
+  --output json --on-error continue
+```
+
+Raise it cautiously: NetBox may rate-limit or strain under high concurrency, so
+tune N to what your instance tolerates. The `--on-error` semantics and the
+one-line-per-record audit log are unchanged. See
+[Writes and safety](writes-and-safety.md#parallel-bulk-writes-workers).
+
 ## Exit-code-driven control flow
 
 ```sh
diff --git a/docs/guides/managing-profiles.md b/docs/guides/managing-profiles.md
index 0956176..247c53e 100644
--- a/docs/guides/managing-profiles.md
+++ b/docs/guides/managing-profiles.md
@@ -96,3 +96,9 @@ NSC_URL=https://other NSC_TOKEN=$TOK nsc dcim devices list   # env vars also wor
 ```
 
 Override precedence (highest first): CLI flag > env var > config profile.
+
+## Shell completion
+
+With shell completion installed, `nsc --profile <TAB>` completes configured
+profile names (read from the cached command-model and config — no network at
+TAB time).
diff --git a/docs/guides/working-with-plugins.md b/docs/guides/working-with-plugins.md
index ab995aa..7cb02b1 100644
--- a/docs/guides/working-with-plugins.md
+++ b/docs/guides/working-with-plugins.md
@@ -30,6 +30,16 @@ nsc commands --schema <path-or-url> -o json | jq '.tags.my_plugin'   # full subt
 document) and emits JSON only. The dumped tree is
 `{info_title, info_version, schema_hash, tags: {<tag>: {resources: ...}}}`.
 
+## Aliases across plugins
+
+The tag-skipping aliases (`ls`, `get`, `rm`, `search`) resolve a resource name
+across **every** tag in the schema, plugin tags included — so
+`nsc ls widgets` finds `my_plugin widgets` without naming the tag, as long as
+the name is unambiguous. Singular forms only apply to the curated core
+resources (`device`, `prefix`, `tenant`, `vlan`, `site`, `rack`, `interface`,
+`cable`, `tag`); plugin resources are matched by their plural name as it
+appears in the schema.
+
 ## Schema cache and plugin upgrades
 
 The cache is keyed by the schema's SHA-256 hash, stored at
diff --git a/docs/guides/writes-and-safety.md b/docs/guides/writes-and-safety.md
index 473087b..103aa3f 100644
--- a/docs/guides/writes-and-safety.md
+++ b/docs/guides/writes-and-safety.md
@@ -47,6 +47,24 @@ nsc dcim devices create -f devices.ndjson --on-error continue --apply   # don't
 `--on-error continue` collects per-record failures, reports them in a summary
 envelope, and exits with the worst error type's code.
 
+### Parallel bulk writes (`--workers`)
+
+When `nsc` loops record-by-record (no bulk endpoint, or `--no-bulk`),
+`--workers N` runs up to N concurrent in-flight requests:
+
+```sh
+nsc dcim devices create -f devices.ndjson --workers 8 --apply
+```
+
+The default is `1` (sequential — unchanged behaviour); the maximum is **32**
+(a higher value is rejected). The per-record `--on-error continue`/`stop`
+semantics are preserved, and the audit log still gets one well-formed JSON line
+per record.
+
+Note that `--on-error stop` bounds only *new* submissions: in-flight requests
+(up to N) still complete. So when the record count is ≤ `--workers`, every
+record is sent regardless of `stop`.
+
 ## Input formats
 
 | Extension | Format |
diff --git a/docs/index.md b/docs/index.md
index 648dce5..1e4ad62 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -8,8 +8,15 @@ because the schema — not hand-written code — defines the surface.
 
 - **Plugins just work.** New endpoints from any installed plugin appear as commands automatically.
 - **Multi-instance.** Named profiles per NetBox instance, plus env-var overrides.
-- **Safe by default.** Writes preview as dry-runs unless you pass `--apply`.
+- **Safe by default.** Writes preview as dry-runs unless you pass `--apply`. An
+  `audit_redaction: full` mode strips every request/response body from the audit
+  log, leaving only `{method, url, status_code, timestamp, profile}`.
+- **Fast bulk writes.** `--workers N` runs up to N concurrent requests on a
+  bulk write (default `1`, max `32`), with per-record error handling intact.
 - **Agent-friendly.** Deterministic command shape, machine-readable JSON output, stable error envelope with documented exit codes.
+- **Curated aliases + completion.** `nsc ls/get/rm/search` over any resource —
+  with singular forms for common ones (`nsc ls device`) — and dynamic shell
+  completion of resource, profile, and enum values from the cached schema.
 - **Interactive TUI.** `nsc tui` opens a keyboard-driven UI to browse, filter, edit, bulk-edit and search — see the [Interactive TUI guide](guides/interactive-tui.md).
 
 ## Three killer examples
diff --git a/skills/netbox-super-cli/SKILL.md b/skills/netbox-super-cli/SKILL.md
index 2832ee2..d49b2c7 100644
--- a/skills/netbox-super-cli/SKILL.md
+++ b/skills/netbox-super-cli/SKILL.md
@@ -15,7 +15,11 @@ endpoint list to go stale.
 - `nsc` is on PATH after `pip install netbox-super-cli`.
 - Config lives at `~/.nsc/config.yaml`.
 - Audit log lives at `~/.nsc/logs/audit.jsonl` (one JSON-line per request/response;
-  passwords are redacted).
+  passwords are redacted). `defaults.audit_redaction` controls how much is
+  kept: `safe` (default) masks sensitive headers/fields but keeps bodies;
+  `full` drops all bodies/headers/query and leaves only
+  `{method, url, status_code, timestamp, profile}` (url stripped of query and
+  `user:pass@`).
 - Cache lives at `~/.nsc/cache/<profile>/<schema-hash>.json`.
 
 ## Command shape
@@ -52,6 +56,11 @@ Plus interactive meta-commands (not aliases):
 - `nsc commands --schema <path-or-url>` — dump the entire generated command
   tree as JSON (useful for discovery; `--schema` is required here).
 
+With shell completion installed, TAB now expands resource names, profiles
+(`nsc --profile <TAB>`), and `--status` enum values. Completion reads the
+on-disk cache only — no network call at TAB time — and silently offers
+nothing if the cache is missing.
+
 ## Dry-run / apply discipline
 
 ALL writes are dry-runs by default. The `--apply` flag is the only path to
@@ -192,6 +201,16 @@ operation.
   ```
 - **Generate the NDJSON once** with whatever scripting tool you prefer
   (`jq`, Python, awk). Don't re-run `list` between every patch.
+- **Speed up large bulk writes with `--workers N`.** Bulk write commands
+  accept `--workers N` (default **1**, max **32**) to keep N requests in
+  flight at once instead of one-at-a-time:
+  ```bash
+  nsc dcim interfaces update -f changes.ndjson --workers 8 --apply
+  ```
+  Per-record `--on-error` semantics are unchanged at any worker count, and
+  the audit log stays intact — appends are serialized so each record is
+  still exactly one well-formed JSON line. Start at the default and raise
+  it only when the record count is large; respect the 32 ceiling.
 
 ### Schema fetches