Simu5G 1.7.0

The most significant change in this release is bearer and QoS management, which now follows the architecture of a real 5G system much more closely: data radio bearers are described by explicit configuration modeled on the spec's own records (TS 38.331 bearer configuration, 5QI characteristics, per-direction QoS-flow classification rules), authored centrally in the new BearerConfigurator module and delivered to the protocol layers that enforce them. This replaces the implicit on-demand creation of bearers, the packet-name heuristics and the scattered per-module parameters of earlier releases, and SDAP is on by default in every network with a 5G core. Dual connectivity gained NE-DC support, SCG bearers and per-bearer split configuration, and the D2D machinery moved out of the core LTE/NR stack into a package of its own, behind a hasD2D switch. Buffer status reporting and uplink scheduling moved to the spec's per-logical-channel-group granularity, which is also what lets the authored QoS profiles reach the uplink scheduler at all. On the channel model, the 3GPP propagation formulas were audited against the reports, fixed, covered with unit tests, and factored into a class family with RAT-neutral names. Changes were validated using fingerprint, statistical and unit tests (details at the end of this entry). This release, like all releases since v1.3.1, was developed by Andras Varga and the OMNeT++ core team.

Tested with INET-4.5.4 and OMNeT++ 6.3, compatible with INET-4.6.0 and OMNeT++ 6.1 through 6.4.

Live Q&A: To discuss this release, I will host an open video call on Wednesday, September 23, 14:00–15:00 CEST (12:00–13:00 UTC) on Google Meet. Anyone is welcome to join, ask questions about the new bearer configuration, migrating existing simulations, or anything else in 1.7.0, or just share feedback. Meeting URL: https://meet.google.com/edr-siof-han

Bearer configuration centralized: the BearerConfigurator module

Every data radio bearer is now described by an entry in the tables of
BearerConfigurator, a new module that exists once per cellular network,
alongside the Binder. It combines, in one network-wide service, decisions
that a real system distributes between the core network and the RAN (in 5GS,
the SMF controls sessions and QoS flows while the NG-RAN maps flows onto
DRBs; in EPS, the roles are spread across the MME and the eNB); the
signaling that would carry the configuration is not modeled, and within the
model the RAN never authors a bearer of its own.

• staticDrbs is the network's bearer configuration: one entry per
  bearer per UE, modeled on the DRB-ToAddMod and RLC-BearerConfig records of
  TS 38.331. A static bearer is configured and established up front, in the
  last initialization stage, so traffic finds it in place. Entries name
  their UE by module path (patterns allowed, so one entry can describe a
  bearer of many UEs), and the configuration follows the UE from cell to
  cell.

• onDemandDrbs describes bearers created when traffic first matches
  them. An on-demand bearer's properties always come from its definition
  entry, never from the packet that triggers it, and a flow that no
  definition covers is a configuration error. The parameter's default value
  keeps unauthored configurations working: it carries the well-known
  packet-name classes (VoIP*, gaming*, VoD*) as ordinary catch-all
  definitions for SDAP-less stacks, and a single header-suppressed default
  DRB for stacks with SDAP.

• An entry states its architecture with the required coreNetwork
  field: a "5gc" bearer is selected by the QFIs mapped onto it
  (mappedQfis; needs SDAP in the stack), an "epc" bearer by packet
  filters (filters, each an inet::PacketFilter -- a message-name pattern
  or an expr(...) expression). Further fields include the QoS profile of
  the bearer's flows (gbr, packetDelayBudget, packetErrorRate,
  qosPriorityLevel -- 5QI characteristics, pushed into the eNB/gNB MAC for
  QoS-aware scheduling), rlcMode and lcg, pduSessionType, isDefault,
  suppressSdapHeader, and the dual-connectivity fields described below.
  See the BearerConfigurator NED documentation for the full schema and the
  modeling abstractions.

• drbProfiles allows named bearer profiles to be defined: a profile
  groups field values (QoS characteristics, rlcMode, and so on) under a
  name, and an entry that references it via its profile field need not
  spell those fields out itself -- the profile supplies defaults for the
  fields the entry does not state. A commonly used subset of the
  standardized QoS characteristics rows is predefined: "qci-1".."qci-9"
  (TS 23.203) and "5qi-1".."5qi-9" (TS 23.501) can be referenced without
  being defined. A row carries what the spec standardizes and nothing else.

• The RLC mode and the logical channel group are RAN choices derived from
  the QoS profile when a definition does not state them: a packet error
  rate at or below amPerThreshold gets RLC AM (a PER target that HARQ
  alone cannot meet gets ARQ), and the priority level is bucketed into an
  LCG by lcgPriorityBounds.

• DRB identities moved to the 3GPP range: drbId is 1..32 (TS 38.331
  DRB-Identity).

• D2D and multicast bearers are outside the definition system: they are
  sidelink bearers, whereas the tables describe infrastructure (Uu)
  bearers. They are established with a fixed configuration (RLC UM, LCG 3).

This replaces every previous bearer-configuration surface, so
configurations that used one of them need updating:

• NrSdap.drbConfig is gone; its role is taken by staticDrbs. An old
  entry like {"drb": 0, "ue": 2049, "qfiList": [1, 2], "rlcType": "UM"}
  becomes {coreNetwork: "5gc", ue: "ue[0]", drbId: 1, mappedQfis: [1, 2], rlcMode: "UM"} -- the UE is named by module path rather than node id, and
  both ends of the bearer are configured from the one entry, so the
  two-sided "rlcType must agree" pitfall of v1.6.0 no longer exists.

• The packet-name traffic classifier is gone. Ip2Nic no longer sorts
  packets into conversational/streaming/interactive/background classes; its
  conversationalRlc, streamingRlc, interactiveRlc and backgroundRlc
  parameters no longer exist, and the LteTrafficClass enum was removed
  (what the scheduler actually consumed all along was the logical channel
  group). The old name-based classes survive as the default onDemandDrbs
  rows, so configurations that relied on the defaults behave as before;
  configurations that set the removed parameters state rlcMode (and
  lcg) on bearer definitions or profiles instead.

• Packet-triggered establishment can be disabled with the new
  establishBearersOnDemand parameter of NrSdap / Ip2Nic, making the
  static configuration the only source of bearers (see the
  VoIP-DL-MultiQfi-NoOnDemand configuration of nr/standalone_drb).

QoS flow classification: the dlQfiRules and ulQfiRules parameters

Which QoS flow (QFI) a packet belongs to is now authored configuration as
well, once per direction at its ingress: BearerConfigurator's dlQfiRules
are delivered to the TrafficFlowFilter at each core-network tunnel entry
(modeling the packet detection rules the SMF installs into a UPF), and
ulQfiRules to each UE's new QosFlowClassifier module (modeling the QoS
rules NAS signaling installs into a UE at PDU session establishment). A rule
matches with an inet::PacketFilter and assigns either a fixed qfi or the
packet's DSCP field read as the QFI (dscpAsQfi); rules are evaluated in
order, first match wins.

Previously the DSCP-as-QFI mapping was hardcoded in TrafficFlowFilter, and
at the UE only reflective QoS could assign an uplink QFI. The dlQfiRules
default value, [{dscpAsQfi: true}], preserves the old downlink behavior. A
UE-classified uplink QFI now survives to the core network, and the new
reflectiveQosOverridesQfi parameter of NrSdap arbitrates between a
classified QFI and a reflective QoS match. NrSdap's useDscpAsQfiFallback
parameter was removed; a {dscpAsQfi: true} rule in ulQfiRules expresses
the same policy, authored where the other classification rules live.

SDAP on by default in 5G standalone networks

hasSdap, which used to default to false everywhere, is now derived from
the network's core: every example network built around a Upf sets
**.cellularNic.hasSdap = default(true) -- the standalone, MEC, cars,
videostreaming and emulation networks all follow. A gNB-served UE in these
networks therefore runs the SDAP sublayer and QFI-based flow-to-DRB mapping,
as a standalone deployment does, and is configured with "5gc" bearer
definitions; the EN-DC networks (an LTE core with NR secondaries) remain
deliberately SDAP-less, matching the architecture they model. Set
hasSdap = false explicitly to keep a 5GC-cored simulation on the old
SDAP-less stack.

Whether a bearer's packets carry the SDAP header on the wire is now a
per-bearer decision (TS 38.331 sdap-HeaderDL/UL), stated with the
suppressSdapHeader definition field and verified per packet at the sender:
suppression is only sound while a single QoS flow rides the bearer, and a
second flow showing up stops the simulation with an error instead of
silently mixing flows. The default (unauthored) DRB is header-suppressed, so
turning SDAP on does not by itself change what the example simulations put
on the air -- their re-recorded fingerprints confirm the packet histories
byte-identical -- but the NIC's module structure changes: an sdap
submodule and a qosFlowClassifier appear.

D2D (sidelink) traffic does not pass through SDAP: SDAP sits between the
core network and the UE, and sidelink flows never touch the core. D2D flows
work identically whether the stack has SDAP or not.

Dual connectivity: NE-DC, SCG bearers, split-bearer configuration

NE-DC -- dual connectivity with an NR master and an LTE secondary -- now
works end to end. The code had NE-DC in its DC role vocabulary but assumed
an LTE master in several places (per-leg id pairing, UE resolution, leg
steering, the X2 mux keying of secondary-leg uplink PDUs); those now read
the actual technologies off the configuration, and a flow's anchor is the
master's cell group whichever technology that is. The NeDualConn example
network moved onto a proper 5GC core, making it the first dual-connectivity
example with SDAP; its new MultiQfi configurations carry SDAP-headered flows
across both legs of a DC bearer.

A bearer definition can now state its dual-connectivity layout:

• legs: the cell groups that serve the bearer -- one of "MCG" /
  "SCG", or both for a split bearer (TS 37.340); a leg element can also
  override the RLC configuration it inherits from the entry. An SCG bearer's
  PDCP still terminates at the master node (an MN-terminated SCG bearer):
  the core network delivers the UE's traffic there, and every PDU crosses to
  the secondary over X2.

• primaryPath, ulDataSplitThreshold, ulLegSelection,
  dlLegSelection: which leg a split bearer's PDUs take. The uplink
  follows the TS 38.323 shape: the UE stays on the primary path until the
  amount queued in its legs' RLC buffers reaches the threshold, then uses
  both legs -- by default the less-loaded one per PDU, or as the
  ulLegSelection expression directs. Steering is by buffer occupancy;
  previously each packet's leg was decided from its IP type-of-service
  marking.

The TechnologyDecision module was removed together with the mechanism it
implemented: IP type-of-service markings no longer influence leg steering
or stack selection. For a non-DC dual-stack UE, the stack a flow uses
follows from the UE's attachment (the useNrCondition parameter, default
typeOfService >= 10, is gone); for DC split bearers, steering is the
per-bearer configuration above. Configurations that set
technologyDecision.typename or relied on ToS-based selection need
updating.

Uplink scheduling and buffer status reporting

Buffer status reporting and uplink scheduling were reworked to the
granularity the spec reports at, the logical channel group. Previously the
eNB/gNB kept one aggregate backlog mirror per UE, the two BSR paths
disagreed about what the UE's backlog is, and the QoS-aware scheduler's
uplink half never saw a bearer's QoS profile at all; a UE with uplink
bearers in more than one logical channel group was scheduled badly, and in
the worst case starved.

• BSRs report per logical channel group (TS 36.321 / TS 38.321 sec
  5.4.5). MacBsr carries a per-LCG size array next to the total, and the
  eNB keeps one backlog mirror per (UE, LCG) -- the new UlBacklogRegistry
  -- instead of one per UE, filed under a pseudo-connection with LCID 0. The
  uplink scheduler's candidate set carries the per-group pseudo-connections,
  so a group is what gets scheduled. MAC control elements occupy no bytes on
  the wire, so the finer report costs nothing there.

• A piggybacked BSR no longer hides the backlog of unscheduled
  connections. The standalone BSR summed all uplink connections, but the
  BSR piggybacked on a data PDU summed only the connections in that TTI's
  schedule list -- so a connection that scheduling passed over never became
  visible to the eNB, which sizes grants from the reported value. Both paths
  now compute the report the same way (each connection's occupancy plus, for
  a connection with backlog, one RLC header of its mode). In the new
  cross-LCG example the lower group's VoIP flows used to deliver 32-37 of
  ~1000 packets over 20 s, at ~10 s mean frame delay, with ~45% of the
  cell's capacity idle -- starvation by invisibility rather than by
  priority; they now deliver all of them.

• An uplink grant is sized from the UE's whole reported backlog. A grant
  is one transport block for the whole UE, which the UE's own LCP then
  divides among its channels; it was sized from the mirror of the single
  group whose pseudo-connection won the scheduling round, so a UE with
  backlog in two groups could never be granted more than the winning group's
  worth in a TTI, however idle the cell. With both of its groups loaded, the
  cross-LCG example now delivers 13.5% more, and its lower-priority flows go
  from 2.3-2.7 s mean frame delay to 33-79 ms.

• QoS-aware uplink scheduling weighs a group by its bearers' profiles.
  QOS_PF's uplink half looked a pseudo-connection's LCID up as a DRB id,
  which no bearer can occupy, so every UE ran at neutral weight (with a
  warning per TTI) and authored QoS profiles never reached uplink scheduling
  at all. A group's weight now comes from the QoS profiles of the DRBs
  established in it, aggregated to the most demanding member on each axis
  (lowest priority level, GBR if any, tightest delay budget and error
  target).

• The priority weight no longer depends on the priority scale.
  QoSAwareScheduler weighed a bearer by 1/(qosPriorityLevel+1), so how
  strongly two bearers discriminate depended on the absolute size of their
  priority numbers rather than their relation: the same pair of services
  weighed 4:3 when authored from the QCI catalog (priorities 2 and 3) and
  31:21 from the 5QI catalog (20 and 30). The weight is the pure reciprocal
  1/qosPriorityLevel now, so only priority ratios matter and the two
  catalogs discriminate identically. qosPriorityLevel is also validated
  against the 3GPP 1..127 range when the definition is parsed: an absent
  (0) or out-of-range level used to rank a bearer silently above every
  legitimate priority. Mixing the two catalogs in one network is still
  wrong, because the absolute values drive the lcgPriorityBounds
  bucketing.

• Connections of the same LCG are served fairly. The UE's LCP served the
  connections of a logical channel group in registration (that is, bearer
  establishment) order, and served each one's entire backlog before looking
  at the next -- so of two backlogged same-LCG bearers the first-registered
  one took the whole grant, and the other starved at trivial load. A group's
  backlogged connections are now served round-robin, one unit per turn (one
  virtual-buffer SDU, or one PDU carve in the NR wire format), and the
  starting position rotates across TTIs so the granularity bias averages out
  instead of always favoring the same connection. LCG index order remains
  strict priority. TS 36.321 / TS 38.321 sec 5.4.3.1: logical channels of
  equal priority should be served equally.

• The QOS_PF proportional quota is computed per node, as its own
  documentation says, but was filed under the node's best-scoring
  connection, so the node's other connections found no quota and were
  granted uncapped. Unreachable while a UE had a single uplink connection,
  routine once uplink backlog is tracked per group.

• The fossil two-phase LCP skeleton inherited from SimuLTE was removed from
  LcgScheduler: an unreachable "switch to best effort" branch, and a token
  bucket computed from hardcoded placeholders that nothing ever read. A real
  prioritized bit rate remains future work -- this removes the pretense of
  one, not the plan for one.

Smaller fixes in the same area: a grant carries the blocks of every codeword
the UE was allocated (dormant today, as uplink allocations are
single-codeword, but it would have gone live with multi-codeword uplink); an
idle D2D connection no longer adds an RLC header to the D2D buffer status
report; the LTE UE's buffer status reports account for RLC header bytes; a
triggered BSR is reported even when the buffers have drained to zero by
reporting time; the standalone BSR the UE main loop schedules is actually
sent; the NR and D2D grant headers are sized in bytes, as plain LTE's are;
and the UE's uplink scheduler clears its scheduled-bytes list each TTI
instead of accumulating an entry per (connection, codeword) for the whole
run. For code that subclasses the MAC: what one connection contributes to a
report is LteMacUe::computeConnectionBacklog(), the rule both reporting
paths draw on, and appendBsr() returns the size it reported instead of the
caller computing it a second time.

These changes move the results of uplink simulations: measurably wherever a
UE has more than one backlogged uplink bearer, and by design where its
bearers span more than one logical channel group. A simulation whose UEs
each have a single backlogged uplink bearer and no sidelink traffic is
unaffected -- keeping those runs byte-identical was the acceptance test each
of these changes was held to. Five new tests in the tests/unit suite pin
the contracts: QosSchedulerWeight, LcgSchedulerFairness,
UeBsrReporting, EnbUlBacklog and EnbGrantFolding.

D2D support factored into a separate package

All device-to-device (D2D) code has been moved out of the core LTE/NR stack
into a dedicated simu5g.stack.d2d package, and D2D is enabled per node via
a single hasD2D switch. The core LTE/NR modules no longer contain any D2D
machinery, and clean (non-D2D) nodes no longer construct it. D2D remains a
research prototype and is not based on any specific 3GPP specification.

• hasD2D node switch. LteUe and eNodeB (and, by inheritance,
  NrUe, gNodeB, LteCar, NrCar) gained a bool hasD2D = default(false) parameter. Setting **.hasD2D = true on a node (or a
  whole fleet) selects the D2D-capable cellularNic variant for that node;
  an explicit cellularNic.typename still works and takes precedence.
  Previously D2D was turned on by naming the D2D NIC type itself
  (*.ue[*].cellularNic.typename = "LteNicUeD2D"), and the NIC's
  d2dCapable parameter told the modules inside it which variant they were
  in. That parameter no longer exists; the node switch has taken over
  its role.

• D2D module types moved to simu5g.stack.d2d, and everything that had
  no D2D type of its own gained one -- D2D used to be baked into the plain
  NrNicUe/NrNicEnb and their submodules on the NR side, and into shared
  modules such as RlcMux, Ip2Nic, LteAmc, Rrc and
  HandoverController on both. The package holds: NICs LteNicUeD2D,
  LteNicEnbD2D and the new NrNicUeD2D, NrNicEnbD2D; MACs
  LteMacUeD2D, LteMacEnbD2D, NrMacUeD2D, NrMacGnbD2D, with the D2D
  AMCs LteAmcD2D/NrAmcD2D and the D2D uplink schedulers; PHYs
  PhyUeD2D, PhyEnbD2D with the D2dChannelModel; RlcMuxD2D and the
  D2D UM entity types; Ip2NicD2D; RrcD2D, HandoverControllerD2D,
  D2DModeController and the mode-selection policies; and the new
  D2dBinder, which holds the global D2D state the Binder used to.
  Module typenames in ini files are unqualified, so these package moves do
  not break existing ini files, and most of these types are selected
  automatically by the D2D NIC anyway.

• The D2D parts of RRC live in RrcD2D. The mode controller and the
  mode-selection submodule moved out of the core Rrc compound into
  RrcD2D, the subclass the D2D NICs select; core Rrc sheds its D2D
  wiring, and with it the hasD2DModeController switch (a D2D RRC always
  has the controller). Switching the periodic mode selection on
  (rrc.hasD2DModeSelection, set by the D2D eNB/gNB NIC) and choosing its
  policy (cellularNic.rrc.d2dModeSelection.typename) work as before, so
  ini files that select a policy are unaffected.

• D2D is a project feature (Simu5G_D2D, enabled by default), so Simu5G
  can be compiled without the D2D code and examples.

Clean NR nodes no longer construct any D2D machinery, so they no longer run
the periodic mode-selection tick or record D2D statistics -- the -nan D2D
scalars that used to appear in non-D2D runs are gone.

Channel model: the 3GPP propagation formulas audited and fixed

The stochastic channel model's path loss, LOS probability, shadowing and
penetration-loss formulas were audited line by line against the reports they
implement (TR 36.814, TR 36.873, TR 38.901), and the defects found were
fixed. Among others:

• LOS probability: UMa used the UMi formula; RMa's and SMa's exponential
  decay constants were swapped; InH lost a 0.54 factor beyond 49 m; and the
  TR 38.901 LOS-probability overrides were never dispatched, so the
  TR 36.873 formulas ran in their place.

• Path loss: RMa and SMa switched to the post-breakpoint slope at the wrong
  distance and evaluated the carrier frequency in the wrong unit; breakpoint
  distances now use the 3.0e8 m/s propagation speed the reports define; the
  TR 38.901 UMa path never drew the 1 m effective environment height and
  missed the tall-UE height draw, and its delegated suburban path aborted,
  or dropped the distance.

• Shadowing: several scenarios chose the shadowing sigma with a breakpoint
  distance inconsistent with the one their path loss used (or with zero), so
  the wrong sigma applied around the breakpoint.

• Building penetration: the O2I model is selected by scenario rather than by
  carrier frequency; TR 36.873's is the flat 20 dB the report specifies;
  TR 38.901's high- and low-loss models were swapped, and its selector
  parameter could never take effect.

Channel state is also keyed correctly now: LOS, shadowing and fading state
belong to a link (a transmitter-receiver pair), not to a node, and the
LOS/shadowing draw is re-anchored each time a link moves a correlation
distance instead of being drawn once per run. D2D receptions are recorded
under the D2D statistics (rcvdSinrD2D) instead of the uplink ones, and the
one-to-many (D2D multicast) reception path runs the same SINR and
reception-decision code as everything else.

These fixes change the statistical results of simulations that use the
affected scenarios and models -- in some cases substantially (a swapped
decay constant or penetration model is a many-dB error). To keep the
formulas fixed, they are now covered by tests: the new tests/unit suite
grades each implementation against oracle values produced by scripts
transcribed verbatim from the reports, and the new simulations/channelmodel
example directory (23 fingerprint configurations) exercises every
propagation formula, both delegation chains, penetration, tall-UE handling,
fading and the sectorial antenna pattern in full simulations.

Path loss formulas factored into a PathLossModel strategy family

The per-3GPP-study propagation formulas (TR 36.814, TR 36.873, TR 38.901),
previously encoded as inheritance depth in the channel-model class chain, now
live in a stateless strategy class family: PathLossModel (abstract) with
concrete Tr36814PathLossModel <- Tr36873PathLossModel <-
Tr38901PathLossModel (the inheritance mirrors each study's own formula
fallback to the previous study, e.g. TR 36.873 has no SMa formulas of its own
and falls back to TR 36.814's). StochasticChannelModel owns one strategy
instance and delegates path loss, LOS probability, shadowing and angular
attenuation to it. Which study to use is selected with the new
pathLossType string parameter ("Tr36814", "Tr36873" or "Tr38901";
default "Tr36814"). Everything else -- fading, interference, SINR assembly,
the reception decision -- is unaffected by the choice of study and stays
shared code.

Channel-model classes and NED types renamed

  LteChannelModel             ->  ChannelModelBase
  LteRealisticChannelModel    ->  StochasticChannelModel
  LteDummyChannelModel        ->  IdealChannelModel
  ILteChannelModel            ->  IChannelModel
  NrChannelModel              ->  Tr36873ChannelModel
  NrChannelModel_3GPP38_901   ->  Tr38901ChannelModel

The old names implied an LTE/NR split which was never actually there -- any
channel model can serve either an LTE or an NR carrier; what varies is which
3GPP propagation study supplies its formulas, and the new names say so.
StochasticChannelModel says how the model works rather than how good it
is: its impairments are drawn from the distributions of a 3GPP propagation
study, as opposed to being computed from the geometry of an actual
environment, and as opposed to the impairment-free IdealChannelModel.

Tr36873ChannelModel and Tr38901ChannelModel are NED-level presets of
StochasticChannelModel (no C++ class of their own) that only override the
pathLossType default, to "Tr36873" and "Tr38901" respectively; they are
now named after the propagation study they select, which is the only thing
that distinguishes them. Both extend StochasticChannelModel directly -- the
former NrChannelModel_3GPP38_901 extends NrChannelModel chain carried no
setting from one preset to the other.

Configurations that name the old NED types explicitly (@class overrides,
ini typename/like selectors, etc.) need to be updated to the new names.

NIC parameter renamed

LteNicBase's lteChannelModelType parameter, which selects the channel-model
NED type plugged into a NIC's channelModel[] submodule vector, is renamed to
channelModelType. NrNicUe's nrChannelModelType, which selects the NR leg
of a dual-leg NIC, keeps its name -- it names a real distinction (the NR leg
of a two-leg NIC), not an accident of the old taxonomy.

Configurations (ini files, NED parameter assignments) that set
lteChannelModelType need to rename it to channelModelType; the old name
is silently ignored rather than rejected, so a configuration using it stops
taking effect without any error being raised.

D2D channel math factored into D2dChannelModel

The device-to-device channel math -- D2D RSRP/SINR computation, D2D
interference, and the D2D reception decision -- used to be built into
LteRealisticChannelModel itself, so every node's channel model carried it
whether the node had D2D or not. It now lives in D2dChannelModel, a
subclass of StochasticChannelModel in the D2D package, which is the
channel model of the D2D NICs (their channelModelType default) on every
propagation study: the inherited pathLossType parameter selects TR
36.814, 36.873 or 38.901 as usual. The d2dInterference parameter and the
rcvdSinrD2D statistic moved with the code; neither exists on the non-D2D
channel models anymore.

D2D configurations normally need not select a channel model at all: the
D2D NICs default to D2dChannelModel, and the study is stated with
pathLossType. The core channel models no longer handle D2D
transmissions.

NrPhyUe removed

NrPhyUe was behaviorally identical to LtePhyUe (the receive path had
long been unified into the base class); it survived only as a marker class
for the dynamic_cast<NrPhyUe *> tests in HandoverController, which told
a dual-stack UE apart from a single-stack one -- for which purpose NrNicUe
gave BOTH of its legs an NrPhyUe (the long-standing "TODO fix this" there).
Those tests now ask the question directly: whether the controller has a
companion-leg otherHandoverController to coordinate with. That leaves
nothing for the marker class to do, so it is removed. Both legs of the
dual-stack UE NICs now run the same PHY types as the single-stack ones.

Configurations that name NrPhyUe explicitly should select PhyUe (see the
rename below) instead.

PHY classes renamed

The PHY module classes are technology-neutral: both the LTE and the NR leg
of every node run the same classes, with per-leg behavior controlled by the
isNr parameter and the channel model plugged into the leg. The Lte
prefix is therefore dropped:

  ILtePhy       ->  IPhy
  LtePhyBase    ->  PhyBase
  LtePhyUe      ->  PhyUe
  LtePhyEnb     ->  PhyEnb
  LtePhyUeD2D   ->  PhyUeD2D
  LtePhyEnbD2D  ->  PhyEnbD2D

Submodule names (phy, nrPhy) and parameters are unchanged, so ini-file
keys are unaffected; only configurations that name the old NED types
explicitly (ini typename selectors, like clauses, @class overrides)
need updating to the new names. The LtePhyFrameType enum keeps its name --
it tags frame types and is not a PHY module class.

Stack opened up for extension

The stack was systematically opened up for external projects that extend
Simu5G by subclassing its modules rather than patching them: some 350
methods across the stack's C++ classes were made virtual (guided by an
explicit rule about what is an extension seam and what is an invariant), the
Rrc compound's submodule types became parametric, the layout of a bearer's
legs is overridable, and per-leg identity (the leg's MacNodeId, its
gates) is resolved from module parameters instead of hardwired assumptions.
None of this changes behavior; the full fingerprint suite is byte-identical.

Other

• Mid-simulation node removal: a node deleted mid-simulation no longer
  leaves state behind that crashes or corrupts the rest of the run: the
  Binder purges all per-node state when a node is unregistered, RRC tears
  down the bearers of a UE deleted mid-run, the AMC forgets D2D feedback
  peers that have left the simulation, and a D2D UM TX entity withdraws
  from the mode controller before it dies.

• Handover: X2-forwarded packets keep their protocol declaration and
  their QFI across the forwarding, and a stale QFI-to-DRB mapping left over
  from before a handover no longer misroutes flows afterwards.

• RRC: dynamic cell association no longer detaches a UE whose
  association scan finds no candidate cell. The serving-node lookup in
  Binder::getServingNode() no longer reads out of bounds.

• PDCP: NrPdcpRxEntity's reorder window handles a full drain correctly
  (a window that emptied could go out of bounds), and the window shift after
  an in-order delivery no longer leaves the last slots stale -- stale slots
  matched unrelated sequence numbers, delivered SDUs from the wrong slot and
  eventually threw a range check. A new test fills the window to capacity.

• Errors that stopped runs part-way are fixed: D2D mode selection
  running before the first D2D CQI report threw std::out_of_range (the
  SinglePair-modeSwitching-TCP configurations of lte/d2d and nr/d2d,
  and FileTransfer-D2D of lte/test_handover, at seed-set 2); the MEC
  real-time video receiver stopped the simulation when a segment of an
  already-skipped frame arrived late, and now discards it as a real-time
  player would (nr/videostreaming_dataset_generator UrbanNetwork); and the
  NrPdcpRxEntity window defect above threw a range check
  (nr/test_numerology MultiCell-CBR-UL, at seed-sets 1 and 2).

• Statistics: the rlcCellThroughputUl / rlcCellThroughputDl
  declarations of RlcMux and macCellThroughputD2D of NrMacGnbD2D are
  removed. The cell-level throughput statistics were dropped in v1.6.0
  because their value was wrong (each cell reported the network-wide total),
  but these declarations came back with the D2D work without an emitter, so
  every .sca file has carried -nan rows for them since. The cell-level
  packet-loss statistics (macCellPacketLossDl / Ul / D2D) are live and
  unchanged.

• GtpUser: locally delivered packets (UE-to-UE within one network) keep
  their QFI.

• D2D multicast: late joiners of a multicast group get their RX leg, a
  remembered multicast flow belongs to its sender rather than just its
  group, and the overlapping multicast group ranges in the nr/cars example
  were fixed.

• Examples: nr/standalone_drb gained MultiQfi and on-demand-bearer
  configurations, lte/tutorial an on-demand-bearer configuration,
  NeDualConn the MultiQfi ones, and a new configuration demonstrates
  uplink QoS classification without the applications' cooperation. For the
  uplink scheduling work, nr/standalone_drb gained VoIP-UL-CrossLcg and
  VoIP-UL-CrossLcg-Heavy (a UE's uplink flows spread over two logical
  channel groups, the second with both groups loaded) and lte/d2d gained
  OneToMany-UDP-D2D (one D2D transmitter with four peers, the scenario in
  which a D2D buffer status report is built from more than one connection).

• Tutorials: tutorials/nr set PacketSize on its CBR senders, a name
  no CbrSender parameter has, so the value was ignored and every tutorial
  configuration ran with the 40 B default -- a 25x lighter load than the
  800 kb/s its comments promise, and than the identical copy under
  simulations/nr/tutorial. The parameter is spelled packetSize now, and
  the two run the same simulation.

• Documentation: the D2D package and its user-guide chapter, and the
  BearerConfigurator NED documentation, which describes the full bearer
  schema and labels its modeling abstractions honestly -- what is spec, what
  is Simu5G policy, and what is not modeled.

Validation

• Fingerprint tests: every change was checked against the fingerprint
  suite, which grew from 157 to 202 configurations in this release, now in
  two CSVs (simulations.csv and simulations_d2d.csv, the latter
  requiring the D2D project feature), covering the channel-model scenarios
  and the new bearer, SDAP, dual-connectivity, cross-LCG uplink and
  multi-peer D2D configurations. Changes meant to preserve behavior had to
  leave the simulated traffic unchanged, as captured by the fingerprints.

• Fingerprints are INET/OMNeT++ version specific: the CSVs hold the
  values for INET-4.5.4 + OMNeT++ 6.3.0, the versions CI uses; other
  versions report mismatches, and they are expected. Over all 202
  configurations: OMNeT++ 6.4.0 moves sz in 181 and nothing else, as it
  writes the mean of an empty statistic as nan rather than -nan;
  INET-4.6.0 moves sz in all of them and tplx/tilx in 18 and 45; and
  INET-4.7.0 is then nearly inert (197 of 202 identical to 4.6.0, the other
  five tplx/tilx only). ~tNl never moves, so the traffic between
  network nodes is unaffected by either version; and no combination
  produced errors. [This section was added post-release]

• Statistical tests: changes that alter behavior were validated by
  evaluating how they move the simulation results. The scalar results of
  the example simulations are kept as baselines in the
  Simu5G-statistics
  repository. They were re-recorded at each stage of development (after
  every result-changing commit for the uplink scheduling work), and each
  deviation was traced to the change that caused it. The repository's
  commit history records the effect of each stage on the results.

• Unit tests: the new tests/unit suite exercises classes directly,
  without running a simulation. The path loss, LOS probability, shadowing,
  penetration loss and antenna pattern of the 3GPP propagation models are
  graded against oracle values computed by a reference script transcribed
  verbatim from the reports. The MAC tests pin the QoS scheduler's weight
  function, the UE's division of an uplink grant among its connections, the
  backlog its buffer status reports announce, and the eNB's per-LCG backlog
  tracking and grant construction.