Adds a 3-stage optimization (sort → merge → packetize) that collapses
scattered read requests into minimal PDU-packed multi-item S7 exchanges.
The optimization is transparent — read_multi_vars() uses it automatically
for PE/PA/MK/DB areas, with CT/TM falling back to individual reads.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…e block splitting

- Parallel dispatch: send multiple packets in-flight on a single TCP
  connection using select(), match responses by sequence number (sliding
  window of max_parallel).
- Auto-tune max_parallel after connect() using PLC's reported
  MaxConnections (via SZL 0x0131), falling back to PDU-size heuristic
  for PLCs that don't support it (LOGO, S7-200).
- Cache optimization plan (sort/merge/packetize) across read_multi_vars()
  calls — recomputed only when the item list changes.
- Fix _split_block to never tear multi-byte items (WORD, DWORD, REAL,
  LREAL) across chunk boundaries.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…ount

The MAX_VARS=20 limit was inherited from the original snap7 C library's
fixed-size array. Our pure Python pipeline sorts, merges, and packetizes
any number of items into PDU-sized packets automatically. The limit is
retained for write_multi_vars where no optimization exists.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Avoids conflict with the upstream python-snap7 package name.
Updates pyproject.toml, publish workflows, and homepage URL.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

python-snap7-optimized was rejected by PyPI name similarity check.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Test jobs don't checkout the repo, so there's no pyproject.toml for cache hashing.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Auto-tune, PDU negotiation, and plan-built messages are now INFO
so they appear in gateway logs without requiring DEBUG on snap7 logger.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

S7-300 PLCs reset the TCP connection when multiple PDUs are sent
back-to-back (parallel dispatch). Force sequential mode for small
PDU PLCs to prevent connection reset by peer errors.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Enable SO_KEEPALIVE with aggressive probes (KEEPIDLE=2s, KEEPINTVL=1s,
  KEEPCNT=2) on all TCP connections for fast dead-peer detection (~4s)
- Add Client.heartbeat_read(timeout_ms) for lightweight liveness checks
  reading 1 byte from M0 with independent short timeout
- Update README with documentation for all optimization features:
  parallel dispatch, plan caching, TCP keepalive, heartbeat read,
  and model-specific tuning table

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

The deploy job fails with 404 because GitHub Pages is not enabled on
this fork. Keep the build job so documentation still gets validated
on push, but remove the upload-pages-artifact and deploy-pages steps.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

parse_get_clock_response was reading year/month/day/hour/min/sec from
bytes 1-6 of the USER_DATA GET_CLK response, but S7-300/400/1200/1500
CPUs put those at bytes 2-7 (bytes 0 and 1 are reserved/version). On
every real PLC the decoder therefore decoded month=26, fell into the
ValueError branch, and silently returned datetime.now() — indistinguish-
able from a real clock read.

Two concrete problems:

1. Wrong offsets: for the observed 10-byte response
     00 19 26 04 17 20 31 55 25 36
   the old code read year=19, month=26 (invalid) → ValueError → fallback.
   The correct decoding is year=26 (2026), month=04, ..., second=55.

2. Silent fallback: both the truncated-response branch and the BCD
   parse-error branch returned datetime.now(). Callers had no way to
   distinguish "this CPU does not expose its clock" (S7-300 answering
   0xd402 "Information function unavailable") from "clock read
   succeeded and happens to equal host time". Gateway tooling using
   this to detect drift would always see ~0ms, including when the PLC's
   real clock was hours off.

Changes:
  - Try year offsets 2 and 1 (handles minor per-CPU variation). Accept
    only results whose full year lands in 1990..2099.
  - Reject responses whose return_code != 0xFF (explicit PLC error).
  - Raise ValueError with diagnostic detail instead of returning
    datetime.now() when the response is truncated or unparsable.

Consumers now get an accurate PLC time or a clear error — no fabricated
fallback values.

Releases the fix in 3cb4751 (parse_get_clock_response correct byte
offsets and no silent datetime.now() fallback). Adds CHANGES.md entries
for 3.1.0 (redundant PLC) and 3.1.1 (this fix).

Single-line the long f-string raises and switch the keepalive
constants to double-quoted strings so pre-commit (ruff-format) is
happy. No behavior change.

heartbeat() already guarded self.connection but not self.connection.socket,
so the settimeout() call tripped mypy's union-attr check. Cache the socket
in a local after the None check and use it for both the gettimeout() read
and the settimeout() write, eliminating the need for additional guards.